Steve Miller's Blog

A digital workshop for systems-minded tech commentary.

Author: steve

  • DHS Ordered DeepSeek Out of Federal Civilian IT and Barred New Purchases: How to Build an AI Model Allowlist Before Your Own Regulator Asks for One

    DHS Ordered DeepSeek Out of Federal Civilian IT and Barred New Purchases: How to Build an AI Model Allowlist Before Your Own Regulator Asks for One

    Banning one app is the easy headline. The hard part, for agencies and companies alike, is answering “where is it actually running?” when models hide inside browser extensions, developer tools, and SaaS features nobody reviewed. That is why the DeepSeek federal ban is worth reading closely, and why it makes a good template for an AI model allowlist of your own. The order is narrower in some ways than the headlines, and broader in one way most people missed.

    What the Order Is, and Who Issued It

    It is not an executive order, an OMB memo, or a new law. It is Exclusion and Removal Order #2026-01, issued by the Secretary of Homeland Security under the Federal Acquisition Supply Chain Security Act of 2018 (FASCSA). The order, published in the Federal Register on October 20, 2026 as DHS Docket No. DHS-2026-0628, is signed by Secretary Markwayne Mullin and copies OMB Director Russell Vought. The notice says the Secretary acted “following review of the Federal Acquisition Security Council Recommendation and affiliated evidence.”

    Dates matter here. The order’s “Original Release Date” is July 24, 2026. Federal Register publication came almost three months later. So the order was already in force well before most people heard about it.

    What It Removes, and From Where

    The order “Applies To: Federal Civilian Executive Branch Agencies except for their national security systems and secure compartmented information systems.” Under FASCSA, the Defense Department and the intelligence community issue their own orders, so this one does not cover them. That is why I retitled the post. “Federal civilian IT” is accurate. “Federal systems,” with no qualifier, is not.

    Within that scope, the “Required Actions” section tells all civilian agencies to:

    1. Exclude “DeepSeek, DeepSeek-affiliated entities, and their products from civilian agency procurements, effective immediately after issuance of this Order,” including “agency-approved products lists, service provider equipment lists, e-commerce markets.”
    2. Remove “products provided by DeepSeek, DeepSeek-affiliated entities, and third-party products and services developed using DeepSeek implementations, from civilian agency information technology systems immediately after the issuance of the Order, or seek an Exception.”

    Two details stand out. First, the phrase “third-party products and services developed using DeepSeek implementations.” This goes beyond DeepSeek’s own app or hosted API. As the law firm Haynes and Boone notes in its analysis, the order “does not define ‘DeepSeek implementation,’” and its removal text “neither distinguishes DeepSeek-hosted from self-hosted deployments nor addresses fine-tuned versions.” In other words, it is unclear whether a vendor’s product that runs a downloaded DeepSeek open-weight model is covered. The text reads broadly enough that a cautious agency would assume it is.

    Second, the stated risks. The order’s risk summary says DeepSeek’s servers are in China and that its account process “collects personally identifiable information and may also track and digitally fingerprint users.” It says prompts may result in “the transfer of sensitive data to the PRC,” and that “the models themselves may contain malware or other security vulnerabilities.” It cites research that found “multiple critical vulnerabilities” in DeepSeek’s iOS app.

    Exceptions, and a Window That Already Closed

    Agencies can ask for an exception, including asking that the order not apply to them, to certain actions, or for a delay. Requests “must be submitted in writing within 60 days of the issuance of this order,” with a “compelling justification” and “any alternative mitigations.” The order names no automatic carve-out for research. Its only categorical carve-out is national security and compartmented systems. Haynes and Boone calculates that, counted from July 24, the 60-day window “closed on Sept. 22, 28 days before publication.” It also notes the order is inconsistent about who reviews requests: one passage says CISA, another says DHS.

    Contractors are reached indirectly. Haynes and Boone explains that the FAR clause 52.204-30 carries FASCSA orders into contracts, requires contractors to check SAM.gov for such orders, and requires a report to the contracting officer within three business days of finding a covered article. That is Haynes and Boone’s reading of the clause, not text from the order itself.

    My Recommendations: Build an AI Model Allowlist

    This section is my opinion and my own playbook, not federal guidance. The order’s structure, exclude from purchasing, remove from systems, document exceptions, is a decent model for any organization.

    1. Inventory first. List every AI app, browser extension, IDE plugin, SaaS “AI feature,” hosted API, and locally run model. Ask vendors in writing which underlying models their features call. Contract language, not marketing pages, is where you will get a real answer.
    2. Write approval criteria. For each model, record who developed it, where it is hosted, where data goes, what the data-retention terms are, and which data classes it may touch. Approve a model per use case, not in general.
    3. Track open-weight provenance. For downloaded models, record the source repository, the exact model ID and hash, and any fine-tunes built on it. The DeepSeek order shows why: “developed using” can reach derivatives, and you will want to answer that question quickly.
    4. Enforce at the network and SaaS layers. Block unapproved AI domains at the secure web gateway or DNS layer, use your CASB or SaaS management tool to disable unapproved AI features, and route sanctioned API traffic through a gateway that logs which model was called.
    5. Control endpoints and developer tools. Use browser and extension allowlists, and watch for local model runtimes and large model-weight downloads on managed devices.
    6. Run an exceptions process with an expiry date. Require a written justification, compensating controls, an owner, and a review date, just as the order asks agencies to do.

    Picture this hypothetical. Your help-desk SaaS ships a new “smart summary” feature, and three months later you learn it runs a fine-tuned open-weight model nobody on your team reviewed. If your allowlist records models per vendor feature, that is a ticket. If it only lists apps, it is a surprise. That is my illustration, not a reported case.

    The Bottom Line

    DHS did not ban DeepSeek from all of government. It ordered federal civilian agencies, outside national security systems, to stop buying it and to remove it, including third-party products built with it, as of July 24, 2026. The open question of what counts as a “DeepSeek implementation” is exactly the question your own allowlist should be able to answer about every model you run.

  • A Forgotten Test Account With the Password ‘Password1’ Exposed Data on 3,300 People, Mostly Children: How to Find the Dormant Accounts Hiding in Your Own Directory

    A Forgotten Test Account With the Password ‘Password1’ Exposed Data on 3,300 People, Mostly Children: How to Find the Dormant Accounts Hiding in Your Own Directory

    Everyone budgets for zero-days. The breach that UK data regulators just turned into a cautionary tale needed no exploit at all. It needed a test account created in January 2012, a password of “Password1,” and two organizations that each thought the other had a handle on directory hygiene. If you want a real-world case study in dormant test account password security, the UK Information Commission’s Office (ICO) just published two of them from the same incident.

    What the ICO Found on Dormant Test Account Password Security

    On October 20, 2026, the ICO announced reprimands to two Welsh organizations following a November 1, 2024 ransomware attack: Ogi Networks Limited (ONL), a managed IT service provider, and the Local Government Data Unit – Wales, known as Data Cymru. The two formal reprimands, dated October 14, 2026, address the same breach from both sides of an outsourcing relationship. Data Cymru acted as a data processor for public-sector projects and retained overall control of its network, including the authority to instruct ONL to create or remove accounts. ONL acted as Data Cymru’s IT service provider and sub-processor since 2012.

    Here is what the primary sources say on the headline claims:

    • The password really was “Password1” on a 2012 test account. According to the ONL reprimand, ONL created an account named “Test2” on January 13, 2012, inside Data Cymru’s Active Directory to test user access before migrating real users. “Test2” was given the password “Password1.” It stayed in Data Cymru’s directory for more than 12.5 years, and had either remained active since 2012 or been inadvertently reactivated before the attack. Checked against Bitwarden’s password-strength tool, the ICO notes, “Password1” is rated “very weak” and can be compromised in less than a second.
    • The victim count is 3,300 people, mostly children. The sheet headline said “thousands of children’s records.” The ICO’s finding is narrower: “During the attack the personal information of 3,300 people was taken, the majority of whom were children,” tied to “two projects supporting children and families.” Because the ICO counts people rather than records and does not give the exact split between children and adults, I retitled the post to match its wording.
    • No ransomware crew is named. Neither the ICO news release nor the two reprimands—nor The Register’s coverage—name a ransomware group or strain. They refer only to “a ransomware attack” and “the threat actor.”

    How a Basic Test Account Became a Full Domain Compromise

    The technical chain in the ONL and Data Cymru reprimands is the part every identity administrator should read twice, because every step is a trap that exists in real production networks.

    First, an incident response firm hired by ONL found suspicious activity on “Test2” as early as September 20, 2024, and on a second test account, “testtest” (created May 8, 2015), on September 24, 2024. Initial access came through a FortiClient SSL VPN connection using “Test2,” likely via password spraying. Neither test account had administrative privileges. They were basic domain user accounts, which was enough to authenticate to an SSL VPN that lacked multi-factor authentication.

    Second, why did a 2012 password survive until 2024? Data Cymru actually had a Group Policy Object (GPO) requiring passwords of at least 15 characters. But as the Data Cymru reprimand notes, “Test2” predated that GPO. Because nobody logged into the dormant account, its password was never changed, so Windows never evaluated it against the newer 15-character policy. Raising minimum password length in Group Policy does not touch existing password hashes in the directory.

    Third, why did account reviews miss it? Data Cymru’s reviews were not scheduled or routine—and when it asked for a user list, it only requested members of the “Domain Users” Active Directory group rather than the full contents of Active Directory. After the ransomware hit, a full audit in November 2024 took just over four hours and found eight deprecated accounts, including “Test2,” “testtest,” and a consultant account with administrative credentials.

    Fourth, once inside, the attacker moved laterally because the local administrator accounts on three affected systems—a primary domain controller, a file server, and an internal SQL server—all shared the same password set in January 2012 and had never been reviewed since. Breaching one system let the attacker collect local account hashes and authenticate across the others. Before deploying ransomware on November 1, 2024, the attacker exfiltrated 42GB of data from those three servers and encrypted 182.7GB across the file and SQL servers.

    Finally, the governance gap was contractual too. Data Cymru lacked a UK GDPR Article 28(4)-compliant sub-processor contract with ONL until April 30, 2025, and had not documented how Active Directory would be monitored or audited. As ICO Head of Investigations David Doodson put it, “The responsibility for these failures did not sit with one organisation alone.”

    My Recommendations: How to Hunt Dormant Accounts in Your Own Directory

    This section is my opinion and my own operational checklist, not official ICO guidance.

    1. Export the whole directory, never a single group. Query every user object across Active Directory or Microsoft Entra ID, regardless of primary group membership, so accounts outside “Domain Users” cannot hide.
    2. Check PasswordLastSet, not just LastLogonDate. An account whose password predates your current length policy is still protected only by your old rule. Reset or disable every account whose PasswordLastSet is older than your current standard.
    3. Put an expiration gate on test and vendor accounts. Flag any account with “test,” “temp,” “vendor,” or “migration” in its name, or with a non-expiring password, and require project closeout checklists to verify removal.
    4. Enforce MFA on every remote access path. Move VPN authentication to SSO backed by Entra ID or your identity provider, enforce MFA via Conditional Access, and audit exclusion groups so test and service accounts cannot log in interactively from outside.
    5. Eliminate shared local admin passwords. Use Windows Local Administrator Password Solution (LAPS) so every server and workstation gets a unique, rotated local admin password, and keep domain controllers isolated.
    6. Put directory audits in the MSP contract. Specify in writing who runs the quarterly full-directory export and who approves every remaining account.

    Picture this hypothetical. Imagine your team raises the domain password policy from eight characters to 16 after an audit. Five years later, an old staging account created for a printer migration—never logged into since—still has its eight-character password and still passes the LDAP filter on a legacy VPN. That is my hypothetical illustration of how policy changes leave legacy holes behind, not a reported customer incident.

    The Bottom Line

    The ICO notes that Data Cymru’s post-incident audit of Active Directory took just over four hours. Skipping that four-hour job for 12.5 years left a “Password1” test account on the perimeter and exposed data on 3,300 people, mostly children. Spend the four hours on your own directory this week.

  • CISA and FBI Say SafePay Ransomware Hit at Least 600 Victims and Wiped Some of Their Backups: Using Advisory AA26-293A to Test Your Own Recovery Before They Do

    CISA and FBI Say SafePay Ransomware Hit at Least 600 Victims and Wiped Some of Their Backups: Using Advisory AA26-293A to Test Your Own Recovery Before They Do

    Most ransomware tabletop exercises end with someone saying, “And then we restore from backup.” The joint SafePay ransomware CISA advisory AA26-293A is a good reason to test that sentence. The agencies say SafePay actors go after recovery before they encrypt anything, and that in some cases they deleted every network-attached backup a victim had. If your recovery plan assumes the backups will be there, this advisory is a free test script.

    What Advisory AA26-293A Says

    The advisory, #StopRansomware: SafePay Ransomware, carries the product ID AA26-293A and an original publication date of October 20, 2026. It is a joint advisory. The authoring agencies are the FBI, CISA, and the U.S. Secret Service, plus Germany’s Landeskriminalamt Rheinland-Pfalz, the UK’s National Crime Agency, and the UK’s North West Regional Organised Crime Unit. It says its TTPs and indicators come from “FBI, USSS, and Landeskriminalamt Rheinland-Pfalz investigations as recently as August 2026.”

    On scale, the advisory says SafePay “was among the most active ransomware groups in 2026” and that its actors “have impacted at least 600 separate victims from multiple critical infrastructure sectors” since August 2024. CISA’s press release of October 21 rounds that to “approximately 600 victims,” and it quotes the FBI’s Brett Leatherman saying the actors “have already targeted more than 600 victims worldwide.” The advisory’s own wording is “at least 600” victims “impacted.” I am using that, not “organizations destroyed.”

    What It Says About Backups

    This is the part that matters most, and the advisory is more specific than “destroyed their backups.” It describes three distinct behaviors, all mapped to MITRE ATT&CK T1490, Inhibit System Recovery:

    • Shadow copies: SafePay actors “halt the Volume Shadow Copy Service.”
    • NAS backups: they “reset Network Attached Storage containing victim backup data to initial factory settings to inhibit system recovery attempts.”
    • Deletion, sometimes: the ATT&CK table adds that “In some cases, SafePay threat actors deleted all network-attached backups of company data.”

    The press release summarizes it as undermining recovery “by deleting or disabling backups.” So the accurate version is this. The advisory describes disabling shadow copies, factory-resetting backup NAS devices, and, in some cases, deleting all network-attached backups. It does not say every victim lost every backup, and it does not describe the backups themselves being encrypted. Encryption is reserved for files, which the advisory says are locked with a ChaCha20 and RSA hybrid scheme and “cannot be recovered without the decryption key.”

    How They Get There

    The advisory’s path to the backups is worth knowing, because each step is a place to catch them:

    • Initial access: credential exposure, brute-force password attacks, phished accounts, and VPN access, including through Mullvad VPN IP addresses. They exploit five listed CVEs in edge devices and internet-exposed systems: CVE-2020-3259, CVE-2023-20269, CVE-2023-27532, CVE-2024-40711, and CVE-2024-40766. Some of these they exploit “through automated scripts.”
    • Persistence and evasion: they take admin rights on existing or new accounts, disable security software, and in some cases change user passwords so valid logins stop working.
    • Lateral movement: PsExec, WinRM, RDP, and remote management tools. They find shares with ShareFinder.ps1 and SharpShares.exe.
    • Exfiltration before encryption: 7-Zip, WinRAR, and FileZilla to collect, WinSCP and RClone to send, and sometimes OneDrive. Victims who don’t pay face leak-site posting and, the advisory says, phone calls from the group.

    My Recommendations: Test Your Recovery Against This Advisory

    This section is my own recommendation. It is built on the advisory’s TTPs and mitigations, but the test plan is mine, not CISA’s.

    1. Ask whether your admin account can wipe your backups. The advisory’s mitigations call for backups that are “encrypted, immutable (i.e., cannot be altered or deleted).” Log in as a domain admin and see whether you can delete a backup set or factory-reset the backup NAS. If you can, so can someone who steals that account.
    2. Check that an offline copy really exists. The advisory says to “Maintain offline backups of data and regularly test and maintain backup and restoration processes.” Find the copy that a NAS factory reset can’t reach. Then prove it is current.
    3. Run a restore without shadow copies. Assume Volume Shadow Copy is gone and restore a real server from your other backups. Time it. That number is your actual recovery time, not the one in the policy document.
    4. Patch the five CVEs, then check whether you were already hit. CISA’s press release urges “urgently patching the known exploited vulnerabilities in this advisory and conducting forensic triage to determine if systems have been compromised.” If any of these products face the internet in your environment, patching alone is not enough.
    5. Separate backup credentials from the domain. The advisory recommends least privilege and just-in-time admin access. I would go further and keep the backup system’s admin accounts out of Active Directory entirely.
    6. Alert on the precursors. Stopping the shadow copy service, mass admin password changes, new admin accounts, and RClone or WinSCP traffic to unfamiliar destinations should all page someone. The advisory’s mitigations include reviewing systems for “new, unactive and/or unrecognized accounts.”
    7. Practice the locked-out scenario. If attackers change your admins’ passwords, how do you get back in? Write the break-glass procedure and test it.

    Here is a hypothetical. Your backup appliance sits on the same domain as everything else, and its admin is a domain group. An attacker who reaches domain admin resets the appliance to factory settings at 2 a.m. and starts encryption at 3. Your restore plan now depends on whatever tape or cloud copy exists outside that domain. That is my illustration of the advisory’s described TTPs, not a reported case.

    The Bottom Line

    The advisory is specific: at least 600 victims, recovery targeted before encryption, shadow copies disabled, backup NAS devices reset, and in some cases all network-attached backups deleted. Read it as a test plan. Run the steps above against your own environment before someone else runs them for you.

  • A Power Event in One Google Cloud London Zone Disrupted Banks and Public Services: Why Your Single-Region Answer Should Be in Writing Before Regulators Ask

    A Power Event in One Google Cloud London Zone Disrupted Banks and Public Services: Why Your Single-Region Answer Should Be in Writing Before Regulators Ask

    Everyone’s architecture diagram says “highly available.” On Tuesday, October 20, a lot of UK organizations found out what that means when one zone in Google’s only London region loses power. The Google Cloud London outage, and the single region risk behind it, is now on the desks of lawmakers and financial regulators. Nobody has asked your company for a written single-region answer yet, but the questions being asked of regulators point that way. Here is what happened, what is confirmed, and what I would write down now.

    What Google Says Happened

    Google’s preliminary incident report, posted October 22, gives the timeline in US/Pacific time. The incident ran from 23:00 on October 19 to 05:10 on October 20, a duration of “6 hours, 10 minutes.” That is about 7:00 a.m. to 1:10 p.m. in London on Tuesday, or 2:00 to 8:10 a.m. ET. Google says the disruption was “triggered by an electrical power interruption at the datacenter facility hosting availability zone europe-west2-b.”

    The root cause, in Google’s words, is a chain of failures. A grid disruption hit the facility. Load moved to UPS batteries while diesel generators started up. But “the Programmable Logic Controller (PLC) automation system (both primary and backup PLCs) failed to initiate the logical automatic start/transfer operations of the generators.” When the batteries ran out, hosts, network switching fabric, and storage in europe-west2-b “lost power and went offline.” Google calls this preliminary and says a final report with preventative actions will follow.

    The part operators should study is what happened next. Google says workloads failed over to the other two zones, europe-west2-a and europe-west2-c, as designed. Then “the sudden influx of redirected query executions, high-availability database failovers, and compute workloads saturated available compute slots and processing capacity in other availability zones.” The report says customers in the unimpacted zones “also experienced unexpected impact,” because of “instantaneous demand surges on unimpacted compute capacity” and “synchronous control-plane validation dependencies across zonal boundaries.” It names BigQuery, Looker, Compute Engine, GKE, Regional Persistent Disk, and Cloud SQL among the affected products. So the zone failed, but the region’s multi-zone design did not fully absorb it.

    Who Was Hit, and Who Says So

    The names come from press reporting, not from Google, which says only that a “very small number” of customers remained affected late in the day. Computer Weekly’s account lists banks and fintechs including Revolut, Lloyds Bank, HSBC, Barclays, Bank of Scotland, NatWest, Halifax, and Nationwide. On the public side it lists NHS England, HM Courts & Tribunals Service, and Gov.uk. Much of that list is based on user reports to DownDetector, and Computer Weekly says the public-sector sites “appeared to be affected.” The clearest confirmations are these:

    • Revolut told customers that “all our customers’ money and data are safe,” Computer Weekly reports. Reuters reports that logging in, card payments, and cash withdrawals were affected, according to Revolut’s website.
    • The Ministry of Justice confirmed that the courts’ Cloud Video Platform had not fully recovered by mid-afternoon, according to Computer Weekly.

    So “froze banks” overstates it. Some banking services at several banks were disrupted for hours. Computer Weekly says most services recovered by late afternoon.

    What Regulators Have Actually Done

    This is where the sheet’s headline went furthest, so I am being careful.

    • UK Parliament, not a regulator, put questions in writing. Computer Weekly reports that Treasury Committee chair Meg Hillier wrote to the Bank of England governor and the FCA’s chief executive. She asked nine questions, with an initial response due by November 4 and full details by January 20, 2027. One question asks: “What lessons should regulated financial firms (rather than Critical Third Parties) take from this incident when thinking about their own operational resilience?”
    • The FCA said it was engaging with Google, according to Reuters.
    • The Bank of England declined to comment to Reuters.
    • In the EU, the European Banking Authority told Reuters it had contacted Google Cloud “to obtain further information on the incident, understand its implications and assess its potential impact on financial entities in the EU.”

    What is already in force: Reuters says Google Cloud falls under both the EU’s critical third-party framework and the UK regime, which designated four cloud providers. Under the UK regime, designated providers “must report serious outages to the FCA and the BoE.” Reuters adds that under UK rules, “banks ultimately are responsible for the services they provide to customers, including when they rely on third-party providers.” What I did not find: any regulator asking firms for a written single-region or concentration-risk answer about this incident. That may come. It has not been reported yet.

    Single Region Risk, in Plain Terms

    Computer Weekly notes that europe-west2 is Google’s only UK region, so customers “that had not built multi-region resilience had no second UK location to failover to.” Google’s own report adds a second lesson: three zones in one region share more than you think. That includes the control plane and the spare capacity everyone fails over into at once.

    My Recommendations: Write the Answer Before Someone Asks

    This section is my opinion, not regulatory guidance.

    1. List every workload pinned to one region. Include the ones that are “multi-zone” but single-region, and the managed services you never configured yourself.
    2. Write down what customers experience if that region degrades for six hours. I mean what they see, not just what your architecture diagram shows. That is the answer a regulator, auditor, or board is most likely to ask for.
    3. Plan for capacity, not just failover. Google’s report shows failover into peer zones can saturate them. Reservations or a second region are how you buy room ahead of time.
    4. Test it. A failover runbook that has never been run is a hypothesis.
    5. Check data residency before you pick a second region. If the only UK region fails, your fallback may be outside the UK. Know whether your contracts and regulators allow that.

    Here is a hypothetical. Your payments API runs regional Cloud SQL across three London zones and your dashboard says “HA.” One zone drops, failover begins, and the surviving zones are already full of everyone else’s failovers. Your status page goes yellow for hours anyway. That is my illustration, built on the failure mode Google describes, not a reported case.

    The Bottom Line

    One London zone lost power because generator automation failed, and the rest of the region struggled to absorb the load. MPs and regulators are now asking questions in writing, mostly of each other and of Google. Your turn may come, so write your single-region answer now, while it is still your own exercise.

  • PHP2Shell: Two Critical PHP RCE Flaws, Made Exploitable With AI Help, Hit the Engine Under Much of the Web: How to Find Every PHP Runtime You Actually Run

    PHP2Shell: Two Critical PHP RCE Flaws, Made Exploitable With AI Help, Hit the Engine Under Much of the Web: How to Find Every PHP Runtime You Actually Run

    On October 21, 2026, the PHP project shipped emergency releases for two critical flaws now widely called PHP2Shell. If you were looking for the PHP2Shell vulnerability patch, it is simple to state: PHP 8.5.12, 8.4.27, 8.3.36, or 8.2.35. The hard part is knowing every place you actually run PHP. This PHP RCE story is less about one server and more about the dozen you forgot.

    What PHP Actually Shipped

    The php.net security advisory says those four releases “fix two critical vulnerabilities that can allow unauthenticated remote code execution,” and that “All PHP users are strongly encouraged to update immediately.” The two bugs are:

    • CVE-2026-109618, RCE via OOM corruption. In php.net’s words, out-of-memory handling “can interrupt PHP while internal request data is only partially initialized. Request cleanup may then treat the invalid data as valid.”
    • CVE-2026-109619, RCE via stale global state. Request-owned internal state “can survive after its backing memory has been freed. A later request handled by the same worker may use that stale state before normal initialization resets it.”

    On scope, php.net says this affects “all PHP installations serving web requests, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed.” It considers the risk “highest for mod_php, with a demonstrated exploit where an unauthenticated request can trigger the attack before any application code runs.” The PHP Foundation’s post adds that remote code execution “was also confirmed on PHP-FPM through application-level code paths.” The PHP security team rates both at CVSS 9.2. The Foundation notes that the CVE IDs leaked a day early with “a wrong CVSS score of 10.0.”

    On older versions, php.net says branches 8.1 and earlier “no longer receive updates from the PHP project.” The security team published “minimal source-backports” for people who build from source. Tenable’s FAQ says those backports cover PHP 7.4 through 8.1. It also cites Calif’s writeup as finding affected code going back to PHP 5.4.0, which gets no fix from the project.

    Was AI Really Involved?

    Yes, by the researchers’ own accounts, though not alone. The PHP Foundation describes a longer history. A Foundation-contracted core developer “found the first signs of the OOM problem” during a 2024 audit, but “no concrete vectors were found.” Then on August 11, 2026, “researcher Ryan A at Calif reported an unconditional remote code execution with the same root cause.” The Foundation credits Calif, Wiz, OpenAI Codex Security, OSTIF, and Quarkslab for research and disclosure.

    Wiz’s research post, which is vendor research, says: “We used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol throughout this research.” It says frontier models enabled them “to develop and adapt working exploits across multiple applications in hours.” Tenable says, “Calif and Wiz have both said they used AI models in their research.” So “found by AI” is shorthand. The underlying problem was spotted by a human in 2024, and AI models helped researchers prove it was exploitable this year.

    What About “Most of the Web” and “15 Minutes”?

    Wiz says PHP “powers 70% of all websites in the world.” Tenable cites a W3Techs figure that PHP is used by “69.3% of websites whose server-side language is known.” That measures how common PHP is, not how many sites are exploitable. Vulnerable versions, server mode, and protections all vary. So “much of the web runs PHP” holds up. “Most of the web is exposed” goes beyond what the sources measured.

    The “15-minute clock” in the original headline traces to a pre-release prediction that mass attacks would start within 15 minutes of the fix. None of the primary sources I read gives that as a measured fact, so I dropped it. What the sources do say about exploitation varies by date and by who is reporting. Tenable quotes Cloudflare’s CTO saying the company is “seeing early attacks emerging,” and says watchTowr reports “exploitation attempts against both.” Those are vendor observations, not counts of breached sites, and I would not assume you have more time than they suggest.

    Find Every PHP Runtime You Actually Run

    This section is my own recommendation as an operator, not guidance from PHP or any vendor. PHP often ships inside other software, so the hard part is finding it. Here is how I would do it.

    1. Distro packages. On each host, check the package manager, not just php -v. A box can have several PHP versions installed side by side. Php.net says distributions and LTS vendors got the patch early, so compare against your vendor’s advisory, since backported fixes may keep an older version number.
    2. CLI vs. FPM vs. mod_php. The CLI binary on your PATH may not be what serves web traffic. Check the FPM service and the Apache module separately. Then restart them. A patched package doesn’t help a worker that is still running the old code.
    3. Containers. Search image registries and running containers for PHP base images and bundled runtimes. Rebuild from patched bases and redeploy. Pulling a new tag does nothing to running pods.
    4. Managed hosts and platforms. For shared hosting, managed WordPress, or PaaS, ask the provider in writing which version you are on and when they patched.
    5. Appliances and bundled apps. Admin panels, monitoring tools, and vendor appliances often embed PHP. These are the hardest to see and the easiest to forget.
    6. End-of-life branches. Anything on 8.1 or older needs an upgrade, a vendor backport, or a decision to take it offline. php.net says these branches get no project updates, only the minimal source backports.

    Here is a hypothetical. Your web fleet is patched by lunch. Then someone remembers the old reporting server in a closet, running a vendor’s dashboard on PHP 7.x behind a forgotten port forward. That is the box your inventory needs to find. It is my illustration, not a reported case.

    If you need time, php.net lists interim settings, such as memory_limit=-1 enforced at the pool level and session.upload_progress.enabled=0. It is blunt that they “reduce exposure to specific attack paths but do not make an installation safe.”

    The Bottom Line

    Two engine-level RCE flaws, confirmed by the PHP project, patched in four supported branches, with exploitation attempts reported by security vendors. AI helped make them exploitable. It did not invent them. Patch, restart, rebuild, and then go hunting for the PHP you didn’t know you had.

  • Moody’s Data Shows Exploited Vulnerabilities Stay Exposed for About Three Months: Closing the Gap Between Your Patch SLA and How Fast Attackers Move

    Moody’s Data Shows Exploited Vulnerabilities Stay Exposed for About Three Months: Closing the Gap Between Your Patch SLA and How Fast Attackers Move

    When a credit ratings firm starts measuring how long you leave the door open, patch speed stops being just an IT dashboard number. It becomes something your CFO, your insurer, and your board can read. Moody’s Ratings has now done exactly that. Its data on the time to patch exploited vulnerabilities puts the typical figure at about three months. Read the method closely, though, because the number is a little different from what the headlines suggest, and the difference matters if you plan to compare it with your own SLA.

    What Moody’s Said, and When

    On October 19, 2026, Moody’s published research arguing that AI-driven cyberattacks are a growing credit risk for companies with weak cyber defenses. I could not read that full report, which is behind Moody’s research paywall. Moody’s own press release, republished by BNamericas as an automatic translation of the São Paulo version dated October 20, says: “The median time to fix a critical vulnerability — or KEV — is about three months, while AI agents that have exceeded the intended scope of controlled test environments have exploited vulnerabilities and compromised systems in a matter of days.” Because that is a machine translation, I treat the exact wording as approximate. The figure itself matches the other sources.

    Infosecurity Magazine’s coverage of the October 19 report gives the specific numbers: known exploited vulnerabilities “take a median time of 87 days to be patched,” compared with 107 days for vulnerabilities with no known exploitation. It says the report “uses February 2026 data from Bitsight Technologies that accounted for vulnerabilities disclosed over previous two years.” It also quotes a caveat from Moody’s: “Meaningful advances in AI’s offensive cyber capabilities since then may have further reduced the average time to exploit; remediation times may also have changed.”

    Where the Three-Month Number Comes From

    The 87-day figure first appeared in an earlier Moody’s report, “Risks posed by unpatched software flaws vary by industry and region,” dated April 1, 2026. That copy is hosted by SecureWorld, not by Moody’s. The method is spelled out there, and it is worth reading slowly. This is ratings-agency data, built on data from Bitsight, which Moody’s calls “our affiliate.”

    • The exact claim: “It takes a typical issuer 107 days for half of non-KEVs to become no longer externally observable, according to our analysis. This compares with 87 days for KEVs and 59 days for ransomware-associated KEVs.”
    • What is measured: not patch deployment. Moody’s says that because the analysis relies on external scanning, “we measure the time from first detection until a vulnerability is no longer externally observable, rather than the precise internal remediation timelines.” A flaw counts as closed “if it is no longer observed for 15 days after its last observed date.”
    • The sample: “around 9,500 issuers,” over two years of data, covering “most rated corporates, financial institutions and a large subset of public infrastructure and public finance issuers.”
    • Coverage: as of February 2026, “Bitsight tracks roughly 35% of KEVs.” So the 87 days describes the KEVs Bitsight can see from the internet, not all of them.
    • The statistic: a median per issuer, meaning the time for half of an issuer’s vulnerabilities to stop being visible. It is not an average and not a worst case.

    That is why I softened the headline. “Typically take three months to patch” is a fair shorthand, and Moody’s own release says “time to fix.” But the underlying measurement is how long a flaw stays visible from outside. A system can stop showing a vulnerability because it was patched, put behind a firewall, or decommissioned. Patching is not the only way out of the data.

    Why the Gap Matters

    Moody’s April report sets the 87 days against two benchmarks. First, “CISA recommends remediation within 21 days for most KEVs.” Second, it cites Flashpoint data that the average time from public disclosure to first exploitation “fell to 44 days in 2025 from 115 in 2024.” It calls that comparison “very conservative,” because Bitsight may take “days or even weeks” to detect a new vulnerability, which “likely understates the true speed advantage of attackers.” It also found that “close to 40% of issuers in any given month had an unpatched KEV that was older than 45 days.”

    The credit angle is in the October press release: “Organizations with poor patching practices will likely fall even further behind,” and organizations with low cyber resilience “will be exposed to higher credit risk.” That is Moody’s forecast, not a finding, but it is a forecast from the people who rate your debt.

    My Recommendations for Closing the Gap

    This section is my opinion, not Moody’s guidance.

    1. Measure your time to patch exploited vulnerabilities the way an outsider would. Your internal ticket says “closed.” An external scan says “still there.” Track both, and look hard at any gap between them.
    2. Make the SLA risk-based, with KEV as its own tier. Generic “critical in 30 days” SLAs treat a CVSS score as the risk. I would put anything in CISA’s KEV catalog on a clock that starts at its listing date, with CISA’s 21 days as the ceiling for internet-facing systems, and treat ransomware-associated KEVs as the fastest lane.
    3. Report medians and tails together. A median of 87 days means half took longer. Moody’s found more than a quarter of issuers had at least one KEV visible for over a year. Your board should see your slowest decile, not just your average.
    4. Count mitigation honestly. Taking a service off the internet is a legitimate fix. Write it down as a mitigation with a review date, so it doesn’t turn into a permanent exception.
    5. Shrink the internet-facing footprint. Moody’s found that 78% of issuers in the top tenth by critical internet-facing systems had old unpatched KEVs, compared with 7.2% in the bottom tenth. Fewer exposed assets means fewer things to chase.

    Here is a hypothetical to make it concrete. Your monthly report shows 95% of critical patches applied within SLA. An outside scan still sees last spring’s VPN flaw on a forgotten secondary appliance. Your dashboard is green while the ratings data could show you as exposed. It is my illustration, not a case from Moody’s.

    The Bottom Line

    Moody’s ratings-agency data says the typical organization takes about 87 days for half of its known exploited vulnerabilities to stop being externally visible, while attackers, by Flashpoint’s count, averaged 44 days to first exploitation in 2025. The number is narrower than “time to patch,” but the gap it shows is real. And now it is measured by someone who isn’t you.

  • Microsoft Calls AI Agents Your Newest Insiders: Give Them an Identity, a Badge, and a Manager Before They Touch Production

    Microsoft Calls AI Agents Your Newest Insiders: Give Them an Identity, a Badge, and a Manager Before They Touch Production

    We spent years building insider-risk programs for employees who might go rogue: access reviews, joiner-mover-leaver processes, and someone watching for the payroll clerk who suddenly downloads the whole HR share. This week, the company selling a lot of us our AI agents said that playbook now applies to software too. That software doesn’t sleep, doesn’t take PTO, and may still hold the admin rights it got in a pilot nobody went back to. Insider threat security for AI agents is no longer a thought experiment. It is a vendor talking point. Before I turn it into a to-do list, it is worth reading what Microsoft actually said.

    AI Agents and Insider Threat Security: What Microsoft Actually Said

    On October 19, 2026, Microsoft published two posts. On the Official Microsoft Blog, Hayete Gallot, Executive Vice President of Microsoft Security, wrote “Securing the agentic insider.” She writes that “frontier models, regardless of their origin, cannot be trusted,” and that “This is not because models are inherently malicious.” Her argument is that models “have been trained to attain results, which often leads to reward hacking.” She says agents “are frequently overprivileged, have access to data, tools and systems and can take actions and make decisions in critical processes.” Then comes the line the coverage picked up: “They are agentic insiders.”

    The same day, Nadim Abdo, Corporate Vice President for Identity and Network Access, published “Agents are your newest insiders. Security must move at their speed.” on the Microsoft Security Blog. His wording is more measured: “With that capability can come insider risk. This does not mean agents are inherently malicious. It means they are capable tools that can be granted access to important systems by users. They can make mistakes, be manipulated, or be compromised.”

    Notice what is missing. Neither post, as I read it, says “insider threat.” Microsoft says “insiders,” “agentic insiders,” and “insider risk,” and both executives say outright that this is not about malice. The “insider threats” framing is TechCrunch’s paraphrase. TechCrunch writes that Satya Nadella has been warning the risk may be agents enterprises deploy, “much like a rogue human employee, otherwise known as an ‘insider threat.’” That is a fair summary of the direction, but it is not Microsoft’s wording. That is why I softened the headline to “newest insiders.”

    The Principle Underneath

    The core idea in Abdo’s post is one any operator will recognize: “surround non-deterministic intelligence with deterministic controls, independent evidence, and containment boundaries. Assume an agent can fail or be compromised.” He quotes Nadella: “The most trustworthy agentic system will not be the one with the model we trust most. It will be the one designed that enables us to trust the model least.”

    This is the same reason we never let an application validate its own audit log. The thing being watched cannot be the thing doing the watching. Abdo puts it as “the model cannot be its own security boundary.”

    What Microsoft Has Documented vs. What It Just Announced

    This is where the Microsoft announcements get blurry, so I am splitting them into two groups.

    Documented Today

    • Microsoft Entra Agent ID. Microsoft’s Entra security for AI overview on Microsoft Learn describes agent identities, agent identity blueprints as templates, an agent registry that lets you “assign identities to agents, autodiscover them across your organization, and manage all agent metadata in one place,” and logging: “All authentication and actions performed by agents are logged in Microsoft Entra ID.”
    • Conditional Access and ID Protection for agents. The same page says Entra applies Conditional Access to agents, including “Microsoft Managed Policies providing a secure baseline by blocking high-risk agents,” and that ID Protection flags “anomalous activities involving agents.”

    Neither Learn page I read carries a preview label on these features. Still, check your tenant and licensing before you plan around them.

    Announced October 19, Availability Not Stated

    Abdo’s post says Microsoft is “introducing” the following. It gives no general-availability dates, so I am treating them as announcements, not shipped products:

    • Dynamic Permissions: task-scoped, time-bound authority, where “Higher-impact actions can require verified human approval.”
    • Dynamic Agent Risk and Dynamic Conditional Access for agents. Abdo describes the response as “graduated, not binary,” for example preserving reads while blocking writes.
    • Guardian Agents and Playbooks in Project Perception.
    • Wider discovery: 35 more local agent types, for “60 supported local agents, including Claude Code, OpenAI Codex, and GitHub Copilot.”
    • New Entra Agent ID identity types for local agents found through sensors. This is an extension of the existing product.

    The post also mentions Microsoft Execution Container (MXC) as a sandbox example, again without availability details. These claims about Microsoft’s own products are Microsoft’s, so treat them as vendor claims.

    My Recommendations: An Identity, a Badge, and a Manager

    This section is my own recommendation, not Microsoft’s guidance, and it applies whatever vendor you use.

    1. Give every agent its own identity. Abdo writes that an agent “should not inherit every permission of the employee who sponsors it.” If your agent signs in as a person or a shared service account, you cannot tell who did what. Fix that first.
    2. Make the badge narrow and short-lived. Start with read-only access to the specific systems the task needs. Add write access deliberately. Put an expiration date on anything granted for a pilot.
    3. Assign a human manager. Every agent needs a named owner who approves its access, answers for its behavior, and gets the alert when it misbehaves. Unowned agents are the “agent sprawl” Microsoft’s Learn page warns about, including agents “created for temporary purposes” that “remain in production indefinitely.”
    4. Run access reviews for agents. Put them in the same quarterly review as people. If nobody can explain why an agent has a permission, remove it.
    5. Write the offboarding runbook now. Know how to disable an agent’s identity, revoke its tokens, and find what it touched before you need to.
    6. Keep monitoring outside the agent. Logs and kill switches should live where the agent cannot change them.

    Here is a hypothetical to make it concrete. A summer pilot gives a reporting agent a global-admin-equivalent role “just to get it working.” The pilot ends, the intern who built it leaves, and the agent keeps running nightly with those rights. Nothing malicious has happened, and nothing has to for that to be a problem. It is my illustration, not a reported case.

    The Bottom Line

    Microsoft did not say your agents are evil. It said they are insiders: capable, often overprivileged, and fallible. I think that framing is right. It is also, as TechCrunch notes, a framing that helps sell Microsoft’s security products. You don’t need to buy anything to act on it, though. Give each agent an identity, a narrow badge, and a manager, and review all three on a schedule.

  • Rank Math PRO Supply Chain Attack Planted a Self-Restoring WordPress Backdoor: How to Vet and Watch the Plugins You Trust

    Rank Math PRO Supply Chain Attack Planted a Self-Restoring WordPress Backdoor: How to Vet and Watch the Plugins You Trust

    “Turn on auto-updates” is the first line of nearly every WordPress hardening guide, and it is good advice. Last week, though, the Rank Math PRO supply chain attack showed why that advice needs a second sentence. A WordPress backdoor arrived inside the official package itself. For about eleven hours, some copies of a paid SEO plugin came from the vendor’s own downloads carrying malware. Keeping everything patched only helps if you also check what the patch actually installed.

    What Rank Math Says Happened

    In its security incident notice, Rank Math says it identified “a security issue affecting a limited number of users who downloaded Rank Math PRO v3.0.123 between October 15, 2026, at 20:02 UTC and October 16, 2026, at 07:12 UTC.” For readers in Eastern time, the notice says that window opened at 16:02 EDT on October 15. The company says, “The affected plugin files contained malware that could compromise your website if the plugin was activated,” and adds, “We immediately removed the affected files and released a clean version, Rank Math PRO v3.0.124.”

    So, to be specific:

    • Edition: Rank Math PRO, the paid plugin.
    • Version: 3.0.123, and only copies obtained during that window. The vendor says “a limited number” of users and gives no count.
    • Fixed version: 3.0.124.
    • The free plugin: not affected. Rank Math says, “This issue affects Rank Math PRO users only. If you use the free version of Rank Math SEO available on WordPress.org, your website is not affected, and no action is needed.”

    The vendor addresses users who “downloaded or updated to the affected version,” so both manual downloads and updates are in scope. BleepingComputer reports that the bad copies came “from the vendor’s official website.”

    What Nobody Has Said Yet: How It Got In

    The headline I was handed said the plugin was “hijacked.” I have not found a source that explains the mechanism. Rank Math’s notice does not say whether its build system, its download servers, or a developer or staff account was compromised. BleepingComputer says it “contacted Rank Math for details on the number of potentially affected sites and the initial attack vector but had received no response by the time of publication.” SecurityWeek calls it a supply chain attack and says “hackers injected malware” into the 3.0.123 packages, but gives no root cause. Until Rank Math publishes one, “hijacked update channel” and “stolen developer account” are both guesses, and I am not going to pick one.

    Does the Backdoor Really Restore Itself?

    According to the vendor, yes. Rank Math warns: “The malware can recreate its hidden backdoor whenever someone visits the website. Removing suspicious files before disabling the mechanism that recreates them may not resolve the issue.” Its manual cleanup section begins with a step titled “Stop the malware from rebuilding itself.” That step means deleting the whole rank-math-pro folder (“Deactivating the plugin alone is not sufficient”), removing a malicious entry from the wp_options table, and blocking outbound connections to rank1cdn.com.

    SecurityWeek’s summary matches: the malware “may recreate a hidden backdoor upon each page visit,” and its persistence “could make manual cleanup a difficult task.” Note the “may.” Both sources describe what the malware can do, not a count of sites where it did.

    Rank Math’s list of indicators to check includes:

    • Administrator accounts with usernames starting “wpsupport_” or email addresses ending in “@wp-mail[.]net.”
    • A fake core file, wp-includes/class-wp-cache-util.php, which Rank Math says “is not a legitimate WordPress core file.”
    • .htaccess rules that let .cache files run as PHP, and .cache files containing PHP.
    • Crontab entries tagged “#ms6v46” for the web server user.
    • Suspicious PHP files in mu-plugins, other plugins, theme folders, and uploads.

    It also says the attacker “may have made changes outside your WordPress installation,” and points to SSH authorized_keys files, other users’ crontabs, and running processes. BleepingComputer adds one finding of its own. It found a PHP sample on VirusTotal that carries the same indicators and includes code to steal cloud credentials and SSH keys. BleepingComputer is careful to say it “could not confirm whether it was distributed through the compromised plugin,” and I will leave it at that.

    If You Run Rank Math PRO

    This part is the vendor’s guidance, condensed. If you installed and activated 3.0.123 during the window, Rank Math says to “Treat your website as compromised.” Its preferred fix is to restore files and database from a backup made before the window, scan, rotate every credential, and then install 3.0.124. It says rotating credentials is “essential, even if you restore your website from a backup.” If you installed but never activated it, delete it, install 3.0.124, and scan. Simply updating to 3.0.124 does not clean an infected site, because the malware’s other files live outside the plugin folder.

    My Recommendations for Vetting and Watching Plugins

    Everything in this section is my own recommendation, not Rank Math’s or anyone else’s.

    1. Stage premium plugin updates. Auto-updates for plugins from WordPress.org are still worth it for most sites. For paid plugins that update from a vendor’s own servers, I would let a staging site take each update first and hold production back by a day. An eleven-hour bad window is exactly what a short delay protects against. The trade-off is that you also delay real security fixes, so keep the delay short.
    2. Keep your own known-good copies. Rank Math notes that checksum verification “may not be available for every plugin, particularly premium plugins.” Save the zip of every premium version you deploy, with a hash. Then you have something to compare against later.
    3. Watch file changes outside update windows. A file-integrity monitor that alerts on new PHP files in uploads, mu-plugins, or wp-includes would have caught several of the indicators above. Run wp core verify-checksums on a schedule, not just after an incident.
    4. Alert on new administrators. New admin accounts on a small site should be rare. Make each one page someone.
    5. Restrict outbound traffic from the web server. Most WordPress sites need to reach only a short list of hosts. An allowlist turns a callback domain into a blocked connection and a log line.
    6. Test your restores. Rank Math’s best recovery path depends on a backup from before October 15. If you have never restored one, you do not really know you have one.

    Here is a hypothetical to show why order matters. An admin finds the fake core file, deletes it, and goes to lunch. The next visitor’s page load rebuilds it. That is the scenario Rank Math’s step order is meant to prevent. It is my illustration, not a reported case.

    The Bottom Line

    What is confirmed: Rank Math PRO 3.0.123, downloaded or updated during an eleven-hour window on October 15 and 16, could carry a backdoor that rebuilds itself. The fix is 3.0.124 plus a real cleanup. The free plugin was not affected. What is not known: how the attackers got in and how many sites were hit. Updates are still the right default. They just are not a substitute for watching what lands on your server.

  • MITRE ATT&CK v20 Adds AI Attack Techniques, MFA Bypass, and CAPTCHA Evasion: Time to Re-Map Your Detection Coverage

    MITRE ATT&CK v20 Adds AI Attack Techniques, MFA Bypass, and CAPTCHA Evasion: Time to Re-Map Your Detection Coverage

    Most teams treat their ATT&CK heat map as a slide they refresh once a year. That habit just got more expensive. MITRE released ATT&CK v20 on October 13, 2026, and the MITRE ATT&CK v20 new techniques for AI abuse, MFA bypass, and CAPTCHA evasion describe the AI assistants and sign-in flows a lot of us deployed in the last year or so. If your coverage map was green last week, some of it is now unmapped, because the rows did not exist yet.

    What MITRE Actually Released

    The official October 2026 release notes list ATT&CK v20 with a start date of October 13, 2026, and call it the “Current version of ATT&CK.” MITRE says “the biggest change in ATT&CK v20 is expanded coverage of recent adversary activity and tradecraft,” including “AI-focused and AI-enabled Techniques in Enterprise and execution and command and control Techniques in ICS.”

    One clarification before the IDs. All of this is in ATT&CK itself, in the Enterprise and ICS matrices. It is not a MITRE ATLAS release. The release notes and changelog I read do not mention ATLAS, so I am not going to guess at how these entries line up with it.

    The Nine New Enterprise Entries, by ID

    The release notes count nine “New Techniques” for Enterprise. Read the detailed v19.2 to v20.0 changelog and that nine breaks down as five new top-level techniques and four new sub-techniques. The difference matters for anyone who maps by ID.

    New top-level techniques:

    • T1699 Direct Prompt Injection. “Adversaries may input malicious prompts directly into an LLM’s user interface.”
    • T1698 Indirect Prompt Injection. Prompts embedded “in content that the adversary expects may be processed by an LLM,” which MITRE says can include web pages, documents, emails, tickets, and code.
    • T1700 AI-Enabled Exfiltration. Using prompt injection “or other methods to instruct enterprise AI agents to exfiltrate data.” This is the data theft piece.
    • T1697 Steal Application Specific Passwords. Stealing these passwords “to gain access to Valid Accounts, bypassing multifactor authentication (MFA) methods.”
    • T1696 Anti-Automation Evasion. The parent technique for getting past CAPTCHA and other bot checks.

    New sub-techniques:

    • T1696.001 Automate CAPTCHA Completion and T1696.002 Avoid CAPTCHA, under the new T1696.
    • T1608.007 Stage Capabilities: AI Result Poisoning, manipulating public web content “to influence generative AI systems or AI-powered search engines.”
    • T1685.007 Disable or Modify Tools: Impair Autonomous Defenses, disrupting “autonomous defensive capabilities such as LLM-assisted processing of malware, artifacts, or activity.”

    Modified, not new: T1195.001 Supply Chain Compromise: Compromise Software Dependencies and Development Tools took a major version bump from 1.3 to 2.0. SecurityWeek reports that it “now also covers IDE and coding agent configuration files.” The changelog also lists a long run of minor version changes to existing techniques, so new IDs are not the only thing that moved.

    A precision note on the headline. “MFA bypass” here means one specific technique, stealing application-specific passwords. It is not a catch-all MFA category. MITRE’s description says those passwords are “typically used to give third-party applications or legacy devices (which do not support MFA methods) account access,” and that an adversary can generate one on an account it already controls.

    The IT/OT Side

    For ICS, the changelog adds T1701 Command and Scripting Interpreter, with sub-techniques T1701.001 PowerShell, T1701.002 Windows Command Shell, T1701.003 Unix Shell, T1701.004 Visual Basic, T1701.005 Python, and T1701.006 JavaScript. It also adds T0869.001 Standard Application Layer Protocol: DNS and T0869.002 Standard Application Layer Protocol: Publish/Subscribe Protocols. The old ICS techniques T0807 Command-Line Interface and T0853 Scripting appear under Deprecations. SecurityWeek says the new technique is “replacing” them.

    Two new campaigns show the IT/OT angle. C0066 2026 U.S. Water Utilities Attacks covers activity against “more than 30 municipal water and wastewater utilities in Minnesota, beginning 27 July 2026.” SecurityWeek notes MITRE’s statement that attribution “remains unconfirmed.” C0064 AI-Assisted Breach of Mexico’s Government Infrastructure is listed in both the Enterprise and ICS domains, which is how MITRE models a cross-domain campaign.

    Why a Green Heat Map Is Now a Stale Cache

    A coverage map is a cached answer to “what can we see?” When the schema underneath it changes, the cache does not throw an error. It just keeps serving the old answer. That is the risk with v20. Nothing in your SIEM broke on October 13. Your dashboard just stopped describing the whole matrix.

    The ICS deprecations are the sneakier issue. If your detection rules, tickets, or vendor reports are tagged T0807 or T0853, those tags now point at retired objects. That will not trip an alert. It will quietly make reporting wrong.

    My Recommendations for Re-Mapping

    This section is my own recommendation, not MITRE guidance.

    1. Diff by ID, not by headline. Pull the machine-readable changelog MITRE publishes with the release and compare it to your detection inventory. Treat the five new techniques, four new sub-techniques, the T1195.001 major version change, and the ICS deprecations as separate work queues.
    2. Mark the new rows honestly. Most teams will not have detections for T1698, T1699, or T1700 on day one. I would show them as gaps, not leave them blank. Blank reads like “not applicable.”
    3. Find your AI agents first. T1700 only matters where an AI agent has tool access to your data. Inventory which assistants can read mail, tickets, or repositories, and what they can send out. Then decide what logging you have.
    4. Audit app passwords. For T1697, check which identity platforms allow application-specific passwords, who has them, and whether you log their creation and use.
    5. Retag ICS content. Move T0807 and T0853 references to the matching T1701 sub-techniques so reporting stays accurate.
    6. Add a purple-team scenario or two. An indirect prompt injection planted in a ticket or document an internal assistant will read is a reasonable tabletop. So is a scripted login flow hitting your CAPTCHA.

    Here is a hypothetical to make it concrete. A support copilot summarizes incoming tickets and can open internal links. An attacker files a ticket with hidden instructions telling it to fetch and post a document externally. Under v20, that chain could map to T1698 and then T1700. It is my illustration, not a case from MITRE’s release.

    The Bottom Line

    ATT&CK v20 is out, and the headline items hold up: two prompt injection techniques, AI-enabled exfiltration, an application-specific password technique that sidesteps MFA, a CAPTCHA evasion technique with two sub-techniques, and new ICS execution and C2 entries. The work is not dramatic. It is a schema migration for your detection program, and like any migration, the risky part is assuming it already ran.

  • Researchers Say One Attacker Hijacked D-Link, Ubiquiti, and SAP Systems as Residential Proxy Exits: When Your Forgotten Device Becomes Someone Else’s Exit Node

    Researchers Say One Attacker Hijacked D-Link, Ubiquiti, and SAP Systems as Residential Proxy Exits: When Your Forgotten Device Becomes Someone Else’s Exit Node

    Search for “residential proxy botnet D-Link Ubiquiti SAP” and you will land on a campaign that GreyNoise published on October 15, 2026. The actual story is narrower than the phrase suggests: one actor, a named set of known vulnerabilities, and at least 143 compromised devices, reported by one security vendor. The sources call it a proxyjacking campaign, not a botnet, and I will stick to their wording. For operators, the interesting part is how little noise it makes.

    What GreyNoise Reported

    GreyNoise, a threat intelligence company, says in its October 15 post that it identified a malicious cyber actor conducting a proxyjacking campaign against network edge devices globally, installing Pawns.app “on behalf of the attacker.” It defines proxyjacking as when an attacker “forces victim devices into proxy networks in order to monetize their bandwidth.” Pawns.app is a legitimate bandwidth-sharing application, which GreyNoise says is a “direct partner” of the proxy provider IPRoyal. The problem is that the attacker installed it without the device owners’ consent.

    GreyNoise says the actor “successfully compromised at least 143 devices in 34 countries,” and lists the exploited vulnerabilities:

    • D-Link ShareCenter NAS: CVE-2024-3272 and CVE-2024-3273.
    • Ubiquiti UniFi Video NVR: CVE-2021-44228, which is the Log4Shell vulnerability.
    • SAP NetWeaver Visual Composer: CVE-2025-31324.

    GreyNoise adds that the actor “appeared to target other brands” of video recorders such as Hikvision and Dahua, without naming the flaws used. In some cases the actor also set up SSH key authentication as “a backup access method.” BleepingComputer’s coverage repeats those details and says GreyNoise found the campaign “while investigating suspicious activity at an unnamed organization.” In GreyNoise’s account, that investigation began when the organization saw residential-looking IP addresses in its sign-in logs, and found multiple users authenticating from the same residential address.

    Does “SAP Boxes” Hold Up?

    Not really, and the difference matters to operators. SAP NetWeaver Visual Composer is enterprise application server software, not a consumer or edge appliance. The U.S. National Vulnerability Database entry lists the affected product as “SAP NetWeaver (Visual Composer development server),” versions “VCFRAMEWORK 7.50,” and describes the flaw as allowing an unauthenticated attacker “to upload potentially malicious executable binaries.” That entry also shows the flaw is in CISA’s Known Exploited Vulnerabilities catalog, added April 29, 2025. Neither GreyNoise nor BleepingComputer says how many of the 143 devices were SAP systems, or which versions. So I will say “SAP systems,” not “SAP boxes,” and I will not suggest SAP environments were a large share of victims.

    The D-Link entry is also specific. The NVD lists CVE-2024-3272 and CVE-2024-3273 as affecting D-Link “DNS-320L, DNS-325, DNS-327L and DNS-340L,” with the note that the issue “only affects products that are no longer supported by the maintainer.” GreyNoise names the product family as ShareCenter NAS rather than listing models, so I am reporting the NVD’s model list as NVD’s, not as GreyNoise’s. Log4Shell is a 2021 flaw that GreyNoise says was used against UniFi Video NVRs, with no model or version given.

    Is It Ongoing?

    The sources do not say. GreyNoise calls it “a recent example” and published indicators, including an IP address it ties to exploitation and the payout account. Neither it nor BleepingComputer gives a start date, an end date, or a statement that the campaign has stopped. So I would treat it as recently observed, with current status unknown.

    For context, the FBI’s March 2026 FLASH on AVrecon malware describes a separate case where routers and other IoT devices were sold as residential proxies through the SocksEscort service, which it says was taken down by law enforcement. It says residential proxies let threat actors “make their activity appear as if it originates from the victim’s network.” That is a different operation and I am not linking the two.

    Why It Matters for Operators

    GreyNoise explains that residential proxies “fundamentally break some of the core assumptions about the detectable characteristics of infrastructure used for malicious activity,” such as “Login from Impossible Travel Location” alerts. BleepingComputer notes that devices in proxy networks can lead to “IP address blocks” or allegations of involvement in illegal activity for the owners, and recalls an earlier FBI warning about such networks being used for credential stuffing.

    My Recommendations for Operators

    This section is my opinion. Where I draw on GreyNoise or the FBI’s FLASH, I say so.

    1. Inventory exposed older gear. Find NAS units, video recorders and appliances reachable from the internet. Anything on a vendor’s end-of-life list is the first candidate. The FBI’s FLASH says that for unsupported devices, “consider replacing the device with a model that is still receiving security updates.”
    2. Check for known-exploited flaws. Compare what you find against the CVEs above. For SAP, confirm whether Visual Composer is installed and exposed, rather than assuming.
    3. Look for unexpected outbound traffic from devices that should be quiet. A NAS or recorder should not be initiating proxy-style connections. Check for unfamiliar SSH keys, since GreyNoise says some victims had them added.
    4. Remember removal is not eviction. If the proxy client is gone but an attacker’s SSH key remains, access remains. I would rebuild or replace a device with confirmed compromise.
    5. Watch for residential IPs in sign-in logs. GreyNoise recommends detecting residential proxy addresses in authentication logs, and especially “multiple enterprise users” signing in from the same one. It also suggests looking at device and session identifiers, not only geolocation.
    6. Restrict egress. Limit what unmanaged or legacy devices can reach outbound, so a hijacked box has a harder time acting as an exit node.

    Picture this hypothetical. An old recorder in a back office has been exposed for years. Nobody pages anyone because nothing is encrypted and no data is stolen. Meanwhile, someone else’s traffic leaves through your address. That is a plausible shape for this kind of abuse, but it is my illustration, not a case from the reports.

    The Bottom Line

    One vendor reported one actor and at least 143 devices, using vulnerabilities that were already public. What we do not know is how widespread this is, which SAP systems were involved, or whether it continues. What we do know is cheap to act on: find your forgotten devices, retire the unsupported ones, and add residential proxy detection to your sign-in monitoring.