Steve Miller's Blog

A digital workshop for systems-minded tech commentary.

Category: Security

  • Vendors and Maintainers Say AI Is Finding Bugs Faster Than They Can Fix Them: Triaging a Vulnerability Flood Without Drowning Your Patch Queue

    Vendors and Maintainers Say AI Is Finding Bugs Faster Than They Can Fix Them: Triaging a Vulnerability Flood Without Drowning Your Patch Queue

    For most of my career, the hard part of vulnerability management was finding the bug. Patching it was tedious, but it was the easy half. That assumption is breaking down. The people running AI-assisted bug hunts are now saying out loud that discovery has outrun the humans who validate and fix what gets found. So the useful skill this year isn’t scanning harder. It’s getting AI-discovered vulnerabilities patch triage right before your queue turns into a landfill.

    I want to be careful with the headline claim. “Faster than teams can fix them” isn’t something I can measure from Mason, Ohio. It’s what the vendors and maintainers doing this work are reporting about their own pipelines. Here’s what they actually said.

    What the People Finding the Bugs Are Reporting

    On Oct. 8, 2026, Anthropic’s Frontier Red Team announced OSS Scanner, a free, opt-in scanning service for open-source projects. The announcement says that over the previous six months its models had turned up “over 29,000 candidate vulnerabilities,” of which the company had been able to “manually review and triage approximately 6,000.” Its own summary: “we remain bottlenecked on our human capacity to validate these findings.” Note the word candidate. Those 29,000 are unvalidated leads, not 29,000 confirmed vulnerabilities.

    Anthropic also says that maintainers increasingly ask for everything in bulk, unverified reports included, and that it has sent “nearly 5,000 reports” that way. The scanner’s output will be “fully model-generated, without human review or triage,” which, by Anthropic’s own wording, means “it is possible reports will be incorrect or invalid.” In plain operator terms, the validation work moves downstream to the maintainer.

    A day earlier, on Oct. 7, Microsoft’s Taesoo Kim, vice president of agentic security, published lessons from Microsoft’s FORGE Lab. Kim writes that from May through September 2026, FORGE helped discover Windows vulnerabilities that were assigned 140 CVEs, 52 of them in the September security release. FORGE members also submitted 155 internally validated reports across 23 open-source projects. The line I’d tape to a monitor: “If reports arrive faster than reviewers … can resolve them, the immediate result is a growing queue.” Kim adds that “a larger search budget can even reduce useful throughput” when duplicates and weak reports crowd out strong ones.

    The maintainer side said it first. In an April 16, 2026 CNCF post, Greg Castle of Kubernetes and Google, writing with contributors from Kubernetes, containerd and Google’s Project Zero, said AI-driven discovery “is already overwhelming OSS maintainers.” He reported that Kubernetes is “dealing with a large volume of vulnerabilities in triage, and starting to get blocked on the next step of developing and releasing fixes.” The post also warns that low-quality AI reports can each “take hours to days to evaluate.”

    Three sources, three vantage points, and the same picture: discovery got cheap, and validation and repair did not. That’s still a set of reports from people inside the work, not an industry-wide measurement. I’d treat it as a strong signal, not a settled statistic.

    Why This Hits Your Queue, Not Just Theirs

    If you run systems rather than maintain libraries, the upstream backlog reaches you as release volume. CNCF expects AI discovery to “create a large volume of patches in the very near term,” and tells projects to expect “more releases than normal” as they and their dependencies ship fixes. Every one of those releases lands in somebody’s change window.

    That’s the correlated-failure problem in a new outfit. Your patch process was sized for a certain arrival rate. If the arrival rate goes up and your staffing doesn’t, the queue grows until something gives. Usually what gives is the change you actually needed, buried under forty you didn’t.

    A Model for AI-Discovered Vulnerabilities Patch Triage

    The following is my opinion, built on the guidance in these sources plus a few decades of watching ticket queues fill up.

    1. Make Reachability the First Gate

    CNCF recommends tools like govulncheck that check whether you actually exercise the vulnerable code, so that “patching that doesn’t carry real security risk” can drop in priority. Microsoft’s post makes the same point from the finder’s side: the bottleneck becomes “determining which reports are real, reachable, and security-relevant.” If you can’t answer “do we even call this code?” quickly, severity scores are just decoration.

    2. Demand Evidence Before Effort

    CNCF says reports without proofs of concept “will be treated as low priority.” Anthropic says each OSS Scanner report includes a self-contained reproducer. Steal that standard internally. A finding with a working reproducer, an explanation, and a version range goes to the front of the line. A finding without them goes back to whoever filed it.

    3. Write Down What Isn’t a Vulnerability

    CNCF’s top recommendation is a published threat model that lists bug classes that commonly get reported but aren’t vulnerabilities. Its example: if the software already requires root, then “taking privileged actions is not a vulnerability.” Your internal version can be a one-page document. It will save more triage hours than any tool you buy this quarter.

    4. Measure the Pipeline, Not the Pile

    Kim’s advice to security leaders is to “count validated fixes, not just findings.” CNCF suggests tracking the number of reports, how many are valid and invalid, the count per severity, and time to triage per report. If all you report upward is “findings discovered,” AI will make your dashboard look heroic while your exposure stays flat.

    5. Set SLAs You Can Actually Hit

    None of these sources prescribes remediation deadlines, so treat the numbers here as an illustrative example, not a standard. Picture a hypothetical shop that commits to 72 hours for reachable, internet-facing criticals with a working proof of concept, 14 days for reachable highs, and the next regular change window for everything else. The specific numbers matter less than the rule underneath them: the clock starts when a finding clears the evidence gate, not when a scanner first emits it. Otherwise you’re running an SLA on noise.

    6. Shrink the Blast Radius Ahead of Time

    CNCF’s closing advice is the least glamorous and probably the most valuable: get off ancient dependencies and old software versions now, so that when a fix lands, “the risk of consuming that patch quickly is much smaller due to a smaller code delta.” A small upgrade is a routine change. A three-major-version jump in the middle of a flood is an outage you scheduled for yourself.

    The Honest Caveats

    Quality is improving, which cuts both ways. Anthropic says expert penetration testers checked 97 critical and high-severity scanner findings across 48 projects, and 85 met its disclosure bar. It also says some maintainers reported inflated severity ratings or a misunderstood threat model. Better reports mean fewer junk tickets, but also more real ones. And these numbers come from vendors describing their own tools, so read them with that in mind.

    My take: the AI part of this story is the scanner. The operations part is everything after it, and that part still runs at human speed. Build your triage so the humans spend their time on what’s real, reachable and exposed, and let everything else wait its turn.

  • Attackers Chained Three Ivanti Connect Secure Flaws Into Unauthenticated Root: Why Your Remote-Access Appliance Needs an Exit Plan, Not Another Hotfix

    Attackers Chained Three Ivanti Connect Secure Flaws Into Unauthenticated Root: Why Your Remote-Access Appliance Needs an Exit Plan, Not Another Hotfix

    The VPN appliance was supposed to be the lock on the front door. After years of emergency advisories, the more useful question isn’t whether this latest patch works. It’s how many more times you’re willing to rebuild trust in the same box. The newest Ivanti Connect Secure vulnerability root access chain makes that question hard to dodge: three flaws, used together in the wild, ending in unauthenticated code execution as root on a device that was fully patched at the time.

    What Volexity Found

    Volexity published its analysis on October 13, 2026. In the first week of October, it found a previously unknown threat actor it tracks as UTA0576 “abusing a zero-day exploit chain to compromise a fully patched (at the time, version 22.8R2.4) Ivanti Connect Secure (ICS) VPN appliance.” In Volexity’s words, the chain “combined an authentication bypass, a remote code execution (RCE), and a privilege escalation vulnerability to achieve unauthenticated root-level code execution.”

    The mechanics, per Volexity:

    • Authentication bypass: a request to the public login script welcome.cgi names a different, protected script in an X-DSCGI-PATH header. The URL decides whether the request is allowed. The header decides where it goes.
    • Code execution: the wu.cgi upload handler writes a file to a predictable path before checking whether it is a valid ZIP. The attacker races to execute it before it is deleted.
    • Root: the root-level dsinvoked service trusts program paths stored in the DSCache configuration store. After patching a library in memory and altering a cached entry, the attacker gets dsinvoked to run a script as root. The exploit then restores the original state and trims log entries.

    Volexity says UTA0576 used the access to extract credentials stored on the appliance, dump limited Active Directory metadata, search Active Directory for LAPS passwords and BitLocker recovery keys, capture LDAP traffic, and run password spraying. It found no backdoor or web shell. The actor re-ran the whole chain for every command. Volexity assesses, with low confidence, that the motivation is criminal rather than espionage.

    What Ivanti’s Advisory Says About the Ivanti Connect Secure Vulnerability Root Access Chain

    Ivanti’s security advisory now covers three CVEs:

    • CVE-2026-104935, 8.6 (High): authentication bypass by spoofing that “allows a remote unauthenticated attacker to execute arbitrary commands.” Volexity notes Ivanti tracks both the bypass and the RCE portion under this CVE.
    • CVE-2026-118527, 4.3 (Medium): use of predictable filenames that lets a remote authenticated attacker upload attacker-controlled content to a known location on disk. Ivanti doesn’t spell out which step this covers; my inference is that it corresponds to the predictable upload path in wu.cgi that Volexity describes.
    • CVE-2026-111357, 8.8 (High): missing authentication for a critical function that lets a local authenticated attacker escalate privileges.

    Ivanti says, “We are aware of exploitation in the wild affecting a limited number of customers at the time of disclosure when CVE-2026-104935, CVE-2026-111357 and CVE-2026-118527 are chained.” So this is confirmed exploitation, not a proof of concept. That said, the public forensic detail comes from Volexity’s investigation of a single customer’s appliance.

    Versions and the Two-Stage Fix

    The fix arrived in two pieces. According to Ivanti’s advisory, CVE-2026-104935 affects 25.1.1.1 and prior, 25.1.2.2 and prior, and 22.8R2.4 and prior. It is resolved in 25.1.1.2, 25.1.3.0 or 25.1.3.1, and 22.8R2.5. Volexity says those patches were released October 13, 2026. The other two flaws affect 25.1.1.2 and prior and 22.8R2.5 and prior, and are resolved in 25.1.1.3 and 22.8R2.6. Volexity’s update says Ivanti published the privilege escalation fix on October 29.

    That sequencing matters. If you patched on October 13 and stopped, Ivanti’s table says you are not on a version that resolves the full chain. Ivanti recommends 22.8 customers move to 25.1.1.3 or later. It notes that 22.8 is meant as “a temporary bridge” and reaches end of life on January 31, 2027. For those who can’t move yet, 22.8R2.6 “resolves all vulnerabilities in this disclosure.”

    CISA’s alert says it added CVE-2026-104935 to the Known Exploited Vulnerabilities catalog on October 13, and the other two on October 30. CISA is “aware of cyber threat actors exploiting these vulnerabilities” to conduct credential theft and data exfiltration.

    Why the Integrity Checker Isn’t Enough

    CISA and Volexity both warn that Ivanti’s Integrity Checker Tool may not detect this. Volexity explains that “because this exploit chain patches library code in memory only and restores configuration store data, the tool does not detect any modifications.” CISA says organizations “will need to conduct network and memory forensics to identify evidence of exploitation.”

    If you suspect compromise, CISA points to Ivanti’s recovery steps. Back up the configuration, factory reset (or rebuild virtual instances) and upgrade, restore the configuration, then reset passwords for all accounts, including domain, administrator, and local accounts. Revoke and reissue stored certificates, and reset stored API keys and license server credentials.

    My Take: Patch Now, Then Plan the Exit

    These are my recommendations, not Ivanti or CISA guidance.

    1. Confirm you are on 25.1.1.3, 22.8R2.6, or later. The October 13 release alone does not cover the full chain.
    2. Hunt using Volexity’s indicators. Look for heavy request volume to welcome.cgi with mismatched X-DSCGI-PATH headers, and check /tmp/. Don’t treat a clean Integrity Checker result as proof.
    3. Rotate what passed through the box. Volexity saw LDAP service account abuse and LAPS and BitLocker key searches. If you can’t rule out exposure, rotate the appliance’s AD service account and review those secrets.
    4. Write an exit plan. 22.8 has an end-of-life date. Use it as a forcing function to evaluate whether an internet-facing VPN appliance is still the right model, or whether identity-aware access with less exposed surface fits better.
    5. Put edge devices in your memory-forensics plan. This attacker built for appliances that don’t log well. Your incident response playbook should assume the same.

    A hypothetical to test your readiness: your appliance was patched October 13, and your Integrity Checker is clean. Can you prove, with network or memory evidence, that nobody ran commands as root on it in the following two weeks? If not, assume the credentials it held are suspect.

    The Bottom Line

    Ivanti, Volexity, and CISA all confirm in-the-wild exploitation of a three-CVE chain that Volexity says reaches unauthenticated root. Patch to the full fix, hunt beyond the Integrity Checker, rotate what the appliance touched, and decide how many more of these cycles you’re willing to run.

  • A WordPress RCE Flaw Drew More Than 18,500 Attack Requests in 21 Hours After the Patch: Shrinking Your Upgrade Window to Hours, Not a Maintenance Night

    A WordPress RCE Flaw Drew More Than 18,500 Attack Requests in 21 Hours After the Patch: Shrinking Your Upgrade Window to Hours, Not a Maintenance Night

    Most of us still treat WordPress core updates as something for the next maintenance night. The October 28 security release is a good reason to rethink that. This is a WordPress RCE vulnerability exploited patch story in the most practical sense. According to Wordfence, attackers began probing for the flaw within hours of the fix shipping and kept at it all night. If your upgrade window is measured in days, the attackers’ window is wider than you think.

    WordPress RCE Vulnerability Exploited Patch: What WordPress Shipped

    WordPress.org published WordPress 7.1.4 on October 28, 2026, describing it as a security release and recommending that sites “update your sites immediately.” The first item on its list is “a weakness in the WordPress upgrade process which can lead to remote code execution during the upgrade window,” reported by Umut Bayram. The release fixes a long list of other issues too, mostly stored cross-site scripting, plus an information disclosure issue and a root certificate update. WordPress says the security fixes were backported to all branches eligible for security fixes, currently through 4.7.

    The Flaw, According to Wordfence

    Wordfence’s analysis, published October 29, lists the issue as CVE-2026-119977. It affects WordPress Core 4.7.0 through 7.1.3 and carries a CVSS rating of 8.1 (High). Patched versions run across every branch, from 4.7.39 up through 6.8.12, 6.9.11, 7.0.8, and 7.1.4. WordPress.org’s own release post does not list CVE identifiers, so the CVE and score here are Wordfence’s.

    The mechanism is unusual. During a Core update that uses a complete WordPress package, affected versions unpack the new files into a predictable directory beneath wp-content/upgrade, and that temporary copy includes the WordPress installer. As Wordfence puts it, “Because the staging directory is web-accessible on many servers, an unauthenticated attacker who requests the installer at precisely the right moment can interact with the staged WordPress copy and progress to PHP code execution.”

    Wordfence is clear that this is hard to pull off. An attacker needs a full-package update, public web and PHP access to the staging tree, a predictable or discovered path, and enough time before cleanup. “In lab environments, the vulnerable staging directory normally remained available for only seconds.” Wordfence says 7.1.4 fixes it by “randomizing the staging directory, adding an index guard, and preventing unsafe promotion of staged configuration data.”

    How Fast the Attempts Started

    Here are Wordfence’s own numbers. Its team deployed a firewall rule on October 28 at 20:30 UTC, “approximately three hours after WordPress announced the security release.” Then: “The first match for our new firewall rule arrived nine minutes after deployment. During the first 21 hours, we recorded more than 18,500 requests targeting staged setup-config.php installer paths.”

    Wordfence describes how the attackers behaved: “Two of the largest sources repeatedly polled small groups of sites for roughly 14.5 hours, behavior consistent with trying to catch the brief upgrade window.” It also adds a fair caveat: “Not every request used the exact vulnerable path.” These are attempts and probes. Wordfence’s post does not report confirmed site takeovers, and neither will I.

    Why does this matter if the window is only seconds? Because the attacker is not betting on one request. Repeated polling is a bet that your update happens while they are watching. The race only exists during an update, which turns the upgrade process itself into the attack surface. Slow, ad hoc, or staggered upgrades give polling attackers more opportunities.

    Shrinking Your Upgrade Window: My Checklist

    These are my recommendations for teams running WordPress at scale, built on WordPress.org and Wordfence guidance. They are not an official checklist.

    1. Verify, do not assume. Wordfence recommends going to Dashboard > Updates and confirming each site is on 7.1.4 or the patched release for its branch. Across a fleet, check versions centrally with WP-CLI or your management tool, and look for sites where auto-updates are disabled or failed.
    2. Let background updates do their job. WordPress says sites that support automatic background updates will begin updating automatically. If you have turned those off for “stability,” treat security releases as an exception with a fast, tested path.
    3. Block web access to the staging directory. The staging tree only matters because it is web-accessible on many servers. Deny public HTTP and PHP execution under wp-content/upgrade at the web server or WAF. I would do this permanently, not just this week.
    4. Put your WAF in front of PHP, not behind WordPress. Wordfence notes this exploit requests a staged PHP file directly without loading WordPress. Its firewall needs “Extended Protection” to see those requests. Whatever WAF you use, make sure it inspects direct PHP requests.
    5. Hunt if you were slow. Wordfence advises reviewing for unexpected requests to PHP files beneath wp-content/upgrade, unfamiliar administrator accounts, unexpected plugins, or PHP files you did not place there.
    6. Set a security-release SLA in hours. Write down that core security releases go out the same day. Measure it, and treat misses like incidents.
    7. Retire unsupported branches. WordPress reminds users that only the most recent version is actively supported, and Wordfence notes 4.6 and earlier get no security updates.

    Consider this hypothetical. An agency manages 200 client sites and pushes core updates manually every Friday night. A security release lands on Wednesday. For two days, unpatched sites carry a known RCE, and every one of those Friday updates opens a fresh race window while polling attackers are already watching. Same-day automated updates plus a blocked staging path close both gaps.

    The Bottom Line

    WordPress shipped the fix on October 28. Wordfence saw the first matching attempts within hours and logged more than 18,500 requests in 21 hours. The exploit is hard, and Wordfence reports attempts, not confirmed compromises. But the speed of the probing is the lesson: your patch window should be measured in hours.

  • HHS Settled Its Ransomware Investigation of Lehigh Valley Health Network for $800,000: What OCR Found About Risk Analysis, Log Review, and Breach Notice

    HHS Settled Its Ransomware Investigation of Lehigh Valley Health Network for $800,000: What OCR Found About Risk Analysis, Log Review, and Breach Notice

    The ransomware crew gets the headlines. The bill that arrives years later comes from the regulator, and it tends to ask a different question: what did you have in place before the attackers showed up? The latest HHS OCR ransomware settlement HIPAA risk analysis case makes that point clearly. On October 29, 2026, the HHS Office for Civil Rights announced it had settled its ransomware investigation of Lehigh Valley Health Network, more than three and a half years after the attack.

    What OCR Announced

    According to the HHS press release, OCR reached a settlement with Lehigh Valley Health Network (LVHN) “and its affiliated covered entities” over potential violations of the HIPAA Privacy, Security, and Breach Notification Rules. LVHN paid $800,000 to OCR and agreed to a three-year corrective action plan. HHS called it OCR’s 22nd ransomware enforcement action.

    The release says LVHN, a Pennsylvania health system, filed a breach report on May 15, 2023, stating that on February 4, 2023 it “discovered suspicious activity in their IT system that was subsequently determined to be ransomware.” Protected health information was exfiltrated and posted on the dark web. OCR puts the number of affected individuals at 248,359, and the exposed information included highly sensitive clinical and identity data.

    Who Was Behind It

    OCR’s release does not name a ransomware group. LVHN did. In a February 22, 2023 statement, President and CEO Brian A. Nester said LVHN “has been the target of a cybersecurity attack by a ransomware gang, known as BlackCat, which has been associated with Russia.” He said the attack hit “the network supporting one physician practice located in Lackawanna County,” that BlackCat demanded a ransom, and that “LVHN refused to pay this criminal enterprise.”

    One date detail is worth noting. LVHN’s 2023 statement says it “detected unauthorized activity” on February 6. OCR’s release says LVHN reported discovering suspicious activity on February 4. I report both as each source states them.

    What OCR Found: HHS OCR Ransomware Settlement HIPAA Risk Analysis Findings

    This is the part that matters for every IT and security leader in healthcare. OCR listed these potential violations:

    • Failing to conduct “an accurate and thorough risk analysis” of the potential risks and vulnerabilities to all of its electronic protected health information (ePHI).
    • Failing to implement procedures to regularly review records of information system activity, “such as audit logs, access reports, and security incident tracking reports.”
    • Impermissibly disclosing the PHI of 248,359 individuals.
    • Failing to timely notify affected individuals of the breach.
    • Failing to timely notify the HHS Secretary of the breach.

    Note the wording. OCR does not say LVHN had no risk analysis. It says LVHN failed to conduct an “accurate and thorough” one covering all of its ePHI. That is a common and important distinction. A spreadsheet that exists is not the same as an analysis that meets the standard.

    OCR Director Paula M. Stannard tied the findings together: “Failure to conduct an accurate and thorough risk analysis and to undertake regular review of information system activity makes regulated entities attractive targets for cyberattacks. The best defense against cyberattacks is full implementation of the HIPAA Security Rule.”

    What the Corrective Action Plan Requires

    Under the plan, which OCR will monitor for three years, LVHN committed to:

    • Conduct an accurate and thorough risk analysis of the potential risks and vulnerabilities to its ePHI.
    • Develop and implement a risk management plan to address the risks that analysis identifies.
    • Review and, as necessary, revise its Privacy, Security, and Breach Notification Rule policies and procedures.
    • Review and, as necessary, develop or revise its HIPAA training materials for workforce members.

    OCR also published general recommendations for regulated entities. They include identifying where ePHI lives and how it flows, putting audit controls in place and regularly reviewing system activity, using authentication so only authorized users access ePHI, encrypting ePHI in transit and at rest when appropriate, and feeding lessons from incidents back into the security management process.

    My Take: Build the Evidence Before You Need It

    These are my recommendations for healthcare IT and security teams and their vendors. They are not OCR guidance or legal advice.

    1. Scope the risk analysis to all ePHI, not just the EHR. LVHN’s own statement says the attack hit the network supporting one physician practice. Affiliated practices, specialty imaging systems, and acquired clinics are exactly where an enterprise risk analysis tends to thin out.
    2. Tie the risk analysis to a living risk management plan. Every identified risk should have an owner, a decision, and a date. OCR’s corrective action plan pairs the two for a reason.
    3. Prove you review logs, not just collect them. Keep records showing who reviewed audit logs and access reports, how often, and what they escalated. “We had a SIEM” is not the same as evidence of regular review.
    4. Rehearse breach notification timing. Two of OCR’s five findings were about late notification. Run a tabletop that starts at discovery and ends with notices filed, and find where your process stalls.
    5. Hold your vendors to the same bar. OCR’s recommendations explicitly cover business associates. Ask vendors for their own risk analysis and log review evidence.
    6. Date and version everything. Investigators look back years. Your documentation should show what you knew and did at the time, not what you reconstructed afterward.

    A hypothetical worth considering: a ransomware group hits a small specialty practice your health system acquired two years ago. Its imaging server never made it into the enterprise risk analysis. When the investigator asks for that analysis, what will you hand over?

    The Bottom Line

    LVHN’s $800,000 settlement and three-year corrective action plan rest on OCR’s findings about risk analysis, log review, and late breach notification. The attackers were BlackCat, according to LVHN. The regulator’s questions were about what was documented and done before and after they arrived. Make sure your answers already exist.

  • Attackers Tried to Exploit a Max-Severity GitLab Flaw Against Alpine Linux’s Own Server: A Patch-Now Checklist for Teams Running Self-Hosted Dev Platforms

    Attackers Tried to Exploit a Max-Severity GitLab Flaw Against Alpine Linux’s Own Server: A Patch-Now Checklist for Teams Running Self-Hosted Dev Platforms

    A self-hosted Git server holds your source code, your CI secrets, and often the credentials that touch everything you ship. So when an open-source Linux distribution’s own GitLab instance becomes the test target, everyone else running one should pay attention. A GitLab vulnerability exploited in the wild means every self-hosted team needs to patch, and the most useful lesson in it is a twist almost nobody expected: Alpine Linux upgraded GitLab and was still vulnerable.

    The Flaw

    GitLab’s October 12, 2026 critical patch release fixed CVE-2026-97229, rated Critical with a CVSS score of 10.0. GitLab describes it as an integer overflow in the CI/CD Regular Expression Compiler that “could have allowed an unauthenticated user to execute arbitrary code on the GitLab server via a specially crafted regular expression in a CI/CD configuration.” Affected versions are 19.2 before 19.2.8, 19.3 before 19.3.5, and 19.4 before 19.4.2. GitLab says GitLab.com and GitLab Dedicated were already patched and urged all self-managed installations to upgrade “immediately.” GitLab also published six threat detections for self-managed customers to spot exploitation attempts.

    On October 29, CISA added CVE-2026-97229 to its Known Exploited Vulnerabilities catalog with a November 1 remediation deadline for federal agencies under Binding Operational Directive 26-04, as BleepingComputer reported. CISA’s entry notes that other products can be affected by the same underlying Ruby regex issue. BleepingComputer also noted that Shadowserver tracks nearly 2,800 internet-exposed GitLab Community Edition instances.

    What Happened at Alpine

    Alpine Linux published its own incident report on October 28, titled “attempted exploitation of CVE-2026-97229.” According to Alpine, on October 20 and 21, “a newly registered account sent crafted CI configurations to our GitLab instance,” which was still running 19.3.4 without the fix. That crashed GitLab worker processes 8 times. “In 4 of those crashes the attacker got partway through the exploit and failed at the next step.” The attacker also downloaded Alpine’s GitLab container image, which Alpine said “suggests they were tuning the exploit to our exact build.”

    Alpine’s bottom line is careful, and I will keep it that way: “We have found no evidence that it succeeded, but we cannot rule it out, so we treated the instance as compromised and rebuilt it from scratch.” Alpine says it found no sign of tampering in its aports repository or releases, and that the keys that sign packages and releases have never been stored on the GitLab server.

    The Twist: Upgrading GitLab Was Not Enough

    Alpine upgraded to GitLab 19.3.6 on October 22. That did not fix the hole. In Alpine’s words, “The fix isn’t in GitLab’s own code but in a patch to Ruby, which GitLab applies only in its official packages. Our image used standard Ruby, so it didn’t get the patch.” GitLab’s release notes do list Ruby regexp patches for each fixed version. Upstream Ruby shipped the fix on October 23, in Ruby 3.3.13, 3.4.12, and 4.0.9.

    Alpine’s guidance for everyone else is direct: “If you run GitLab with your own Ruby (from source or in your own container image) rather than GitLab’s packages, upgrading GitLab is not enough: you also need Ruby 3.3.13, 3.4.12, 4.0.9 or later.” Alpine also admits that on October 13 it upgraded to 19.3.4 instead of 19.3.5 “by mistake.” I respect that candor. It is exactly the kind of slip that happens on any busy team.

    Why Alpine Could Not Be Sure

    This part should make every operator check their logging. Alpine says a successful exploit “would leave no crash and no trace in the web logs.” Some logs from the attack day were gone, and “we had no logging of outbound network connections, and our hosting provider keeps none either.” If the attack had worked, Alpine lists what could have been exposed, including password hashes, 2FA secrets, CI/CD variables and runner tokens, webhook and mirror credentials, and private repositories.

    A Self-Hosted Patch Checklist for the Exploited GitLab Vulnerability

    These are my recommendations, built on GitLab’s and Alpine’s guidance. They are not an official checklist from either.

    1. Confirm your version against GitLab’s advisory. You need 19.2.8, 19.3.5, 19.4.2, or later. Double-check the exact number. Alpine’s 19.3.4-versus-19.3.5 slip is easy to repeat.
    2. Check your Ruby if you do not use GitLab’s official packages. Source installs and custom container images need Ruby 3.3.13, 3.4.12, 4.0.9, or later, per Alpine. Verify the Ruby version inside the running container, not just your build file.
    3. Deploy GitLab’s published detections. Worker crashes tied to the CI regex compiler are your early warning. Alpine’s first signal was crashes.
    4. Review recent sign-ups and CI lint activity. Alpine’s attacker used a newly registered account. If open registration is enabled on an internet-facing instance, ask whether it needs to be.
    5. Start logging outbound connections from the GitLab host now. Without egress logs, you cannot prove a negative. Keep logs longer than two weeks. Alpine says it is moving past 14 days.
    6. If you were exposed and cannot rule out compromise, rotate. Runner tokens, CI/CD variables, personal access tokens, webhook and mirror credentials, and user passwords. Alpine reset all passwords and replaced GitLab’s secret keys.
    7. Keep signing keys off the Git server. Alpine’s best decision was made long before this incident.
    8. Reduce exposure. If your instance does not need to be public, put it behind a VPN or identity-aware proxy.

    Picture this hypothetical: your team patched GitLab last week and closed the ticket. Your instance runs in a custom image on stock Ruby. By Alpine’s account, you may still be exposed, and your web logs would not tell you.

    The Bottom Line

    CISA now lists this CVSS 10.0 flaw as exploited, and Alpine’s report shows an attacker came close against a well-run open-source project. Alpine found no evidence of success but could not rule it out. Patch GitLab, verify Ruby, turn on egress logging, and decide now what you would rotate if you could not prove a negative.

  • The Pentagon Is Treating Freezer Malfunctions on Military Bases as a Suspected Hack: Who in Your Organization Owns the Building Systems on Your Network?

    The Pentagon Is Treating Freezer Malfunctions on Military Bases as a Suspected Hack: Who in Your Organization Owns the Building Systems on Your Network?

    When a freezer fails, the first call usually goes to facilities, not the security team. That habit is about to get tested. Reuters reported on October 29, 2026 that the Pentagon is treating a string of freezer malfunctions at U.S. military bases as a suspected hack. That puts building management system cybersecurity, and the refrigeration and facility controllers that sit next to it, squarely on the agenda for anyone who runs a network with equipment from a facilities vendor plugged into it. Which is most of us.

    What Has Been Reported

    According to Reuters, as published by Military Times, the Pentagon is treating “the spate of mysterious freezer malfunctions that spoiled food at military bases across the United States in August as a suspected hack,” according to four officials familiar with the matter. Reuters said it could not establish how far the investigation had progressed or whether it had reached any conclusions about what happened or who was responsible. The Pentagon said a unit of Cyber Command was “working in close coordination with interagency and law enforcement partners.” Army CID said its Cyber Field Office was involved and was working with French law enforcement on “the refrigeration outages in Europe.”

    Reuters’ review of the bases’ own Facebook posts and website updates found that the late-August malfunctions affected at least six commissaries, the on-base grocery stores. In at least one case, the freezers did not simply stop. Their temperatures were turned up overnight. Fort Huachuca, Arizona posted that its freezer “went into defrost mode where it heated up all the food.”

    Back on September 1, Stars and Stripes reported that Army CID’s Cyber Field Office and the Air Force Office of Special Investigations had confirmed active investigations. Based on each installation’s online announcements, it named the affected stores as Fort Irwin, F.E. Warren Air Force Base, Fort Huachuca, Naval Station Newport, Columbus Air Force Base, and Travis Air Force Base. At that point, the Defense Department would acknowledge only a “possible refrigeration disruption.”

    On attribution, Nextgov/FCW reported that U.S. officials suspect Iran-linked hackers, citing two anonymous people familiar with the matter. Nextgov also said it is not clear how the intruders may have accessed the freezer systems, and that the assessments could change. I want to be precise here. As of these reports, no official has publicly confirmed that this was a cyberattack, how it was done, or who did it.

    Why the French Parallel Matters

    Reuters connects the U.S. investigation to a cyberattack earlier in October on commercial freezers in France. That attack overrode temperature controls at more than 350 establishments and, like Fort Huachuca, involved raising temperatures rather than just shutting units off. Danfoss, whose controllers ran many of the affected French units, said many devices were obsolete or running outdated firmware. Reuters could not establish whether Danfoss or any other controller brand was in use at the U.S. bases.

    Bill Moore, CEO of Xona Systems, told Reuters the root problem is that controllers “initially conceived as isolated interfaces that would be serviced by an in-person technician were increasingly being connected to the internet.” That one sentence describes a lot of the building systems I have seen on corporate networks.

    The Ownership Gap in Building Management System Cybersecurity

    Here is my opinion: building systems fall through organizational cracks. Facilities buys the HVAC controls, refrigeration, access control, and lighting. A vendor installs them, often with remote access for service. IT provides a network port, or the vendor brings a cellular modem. Security never sees an asset inventory. When something goes wrong, the ticket says “equipment failure,” and nobody asks whether someone changed a setpoint on purpose.

    The incidents in these reports show why that matters. A freezer that heats up overnight looks exactly like a mechanical fault until someone checks who, or what, changed the configuration.

    My Recommendations: Assign the Owner Before the Incident

    These are my recommendations, not Defense Department or vendor guidance.

    1. Name a single accountable owner for building and facility systems security. This is a governance decision, not a technical one. Write it down. Facilities can operate the systems, but someone in security must own their risk.
    2. Inventory every controller. HVAC, refrigeration, BMS head-ends, access control, elevators, metering. Record the vendor, firmware version, network path, and how the vendor connects for remote service.
    3. Find anything internet-exposed. Check your external attack surface for building system interfaces. Assume an exposed controller with old firmware will be found by someone else first.
    4. Segment and broker vendor access. Put building systems on their own network segments. Route vendor remote access through a monitored, MFA-protected path, not an always-on modem or shared password.
    5. Log and alert on setpoint changes. A temperature target changing at 2 a.m. is a security event until proven otherwise.
    6. Add a “suspected tampering” step to facilities runbooks. Before resetting a failed unit, capture configurations and logs. Preserving evidence costs minutes. Resetting first can erase the only clue.
    7. Patch or replace obsolete controllers. Danfoss’s own explanation of the French incident is about outdated devices. Put end-of-life controllers on the capital plan.

    A hypothetical to test your plan: your pharmacy, lab, or cafeteria cold storage alarm goes off overnight. Does the on-call facilities tech know to call security? Does security know where the controller sits on the network? If either answer is “probably not,” you have found your gap.

    The Bottom Line

    The Pentagon investigation is reported, not concluded. Officials have not publicly confirmed an attack or named a culprit. You do not need a verdict to act on the lesson. If equipment on your network can change physical conditions, someone in your organization needs to own its security, and everyone needs to know who that is.

  • Your AI Agents Are Holding the Keys: Why Their Credentials Are Becoming a Critical Security Boundary

    Your AI Agents Are Holding the Keys: Why Their Credentials Are Becoming a Critical Security Boundary

    The fastest way to give an AI agent access to something is to paste an API key into its config. That is how a lot of pilots started, including some I have watched up close. It works right up until the agent gets prompt-injected, logs the key, or sends it along to a model provider with the rest of its context. Managing AI agent secrets and credentials is moving from a niche concern into product announcements from both a fast-growing startup and IBM. If your team runs agents in production, the shortcut you took last quarter is worth revisiting this week.

    Two Vendor Moves in the Same Week

    On October 27, 2026, two very different companies made agent credentials the center of their announcements.

    Infisical, the open-source secrets management company, announced a $40 million Series B at a $600 million valuation, led by Elad Gil. The release describes its Agent Vault product this way: it “lets companies give AI agents access without handing over credentials.” CEO Vlad Matsiiako put the logic plainly: “Agents need access to company systems to do their jobs, but giving them actual passwords and API keys is a major security risk.” In his own post on the round, Matsiiako wrote that Infisical will “significantly increase our focus on Agent Vault.” He also wrote that the company “didn’t have to raise this round” because it has been cash-flow positive. Infisical says it serves more than 1,300 customers.

    IBM, at its TechXchange 2026 conference, announced that HashiCorp Vault “will now extend those capabilities to AI agents, issuing secured, short-lived credentials that limit each agent to what it needs.” IBM paired that with Agent Identity, which it describes as establishing “the authoritative record for each agent, such as who owns it, its lifecycle status and the authority it has been granted.” IBM says HashiCorp Vault 2.2 and Agent Identity are generally available. It also says Vault 2.2 “gives agents their own secured accounts” and automatically renews their credentials.

    Is money “pouring into” this space? I can verify one funding round here, Infisical’s, plus a large incumbent repositioning its products. That is a meaningful signal. It is not proof of a market-wide flood, and I am not going to claim one.

    Why Credentials Are the Pressure Point

    IBM’s release makes the argument better than I could: “As agents invoke tools and act across business systems, their credentials become a critical security boundary.” It continues: “An agent with a long-lived, overprivileged credential can put systems and data at risk if it’s compromised or acts outside its intended task.”

    Infisical’s release frames it more bluntly. Agents need credentials like API keys, passwords, and authentication tokens, “which AI agents cannot be trusted with.” The company says Agent Vault is built so an agent “cannot leak credentials in a cyberattack or inadvertently send them to LLM providers or other tools it communicates with.”

    Here is my opinion, and I am labeling it as one. For many teams, secrets management is now the front line of agent security. That is because it is the control you can actually enforce today, while prompt-injection defenses are still immature. You cannot fully stop an agent from being tricked. You can limit what a tricked agent is holding.

    AI Agent Secrets Management: Two Architectures, One Principle

    The two announcements take somewhat different routes to the same idea:

    • Brokered access (Infisical’s description): the agent never holds the real secret. Infisical says it lets agents reach approved systems without exposing the underlying credentials.
    • Short-lived, identity-bound credentials (IBM’s description): each agent gets a governed identity and credentials that are short-lived and least-privilege. IBM says access can be “denied and session revoked in real time” when an agent tries something outside its approved permissions.

    Both reject the pattern most pilots started with: one long-lived, broadly scoped static key, sitting in a config file or environment variable, with no owner and no expiry.

    What I Would Do on Monday

    These are my recommendations for platform and security teams, not vendor guidance.

    1. Find the keys you already handed out. Scan agent configs, prompt templates, notebooks, CI variables, and repos for hardcoded secrets. IBM’s release mentions customers using HashiCorp Vault Radar to detect hardcoded and exposed secrets. Whatever tool you use, inventory first.
    2. Give every agent an owner and an identity. No shared service accounts across agents. If you cannot name who owns an agent, you cannot decide when to revoke its access.
    3. Replace static keys with short-lived, scoped credentials. Issue them per task or per session, scope them to the specific systems and actions the agent needs, and let them expire on their own.
    4. Keep secrets out of the model’s context where you can. A proxy or broker pattern means a prompt-injected agent has nothing reusable to exfiltrate. Remember that it can still leak the data an approved call returns.
    5. Log every credentialed call. Audit trails should tie each action back to an agent identity and a human owner.
    6. Rehearse revocation. Practice killing a misbehaving agent’s access in minutes, not hours.

    Try this hypothetical. A coding agent with a long-lived cloud key reads a poisoned README that tells it to print its environment variables into a pull request comment. With a static key, you are rotating credentials and reviewing logs all weekend. With a brokered or short-lived credential, the leaked value is useless or already expired. That difference is the whole argument.

    The Bottom Line

    Infisical’s Series B and IBM’s Vault and Agent Identity launches both point at the same weak spot: agents holding credentials they should not keep. The tooling is arriving. The work that falls to your team is finding the keys already pasted into configs, then moving every agent to owned, scoped, short-lived, audited access.

  • CISA 2030 Promises an Aggressive, Anticipatory Stance to Eliminate Adversarial Footholds in Critical Infrastructure: What the Strategy Actually Asks of the Operators Who Run It

    CISA 2030 Promises an Aggressive, Anticipatory Stance to Eliminate Adversarial Footholds in Critical Infrastructure: What the Strategy Actually Asks of the Operators Who Run It

    CISA released its new strategy, CISA 2030: Defend, Harden, Prevail, on October 26, 2026. The headline promise is bold. In his welcome letter, Acting Director Nick Andersen writes that CISA “will transition from passive defense to an aggressive, anticipatory stance that actively eliminates adversarial footholds in defense-critical infrastructure prior to weaponization against our systems.” If you run a water plant, a regional utility, a hospital network, or any other operational environment and you are reading up on the CISA strategic plan, critical infrastructure, and proactive defense, the question that matters is practical. What does this plan actually ask of the operators? I read the whole document. The honest answer is: less than the rhetoric suggests, at least so far.

    What the Strategy Says

    The document is short, about eight pages of substance. It is organized around three pillars, each named in the plan:

    • Defend: Lead National Cyber Defense
    • Harden: Prioritize National Infrastructure Security
    • Prevail: Mobilize National Resilience

    Andersen frames the premise realistically. He writes that “there is no guarantee that a strong defense will eliminate intrusions,” and that “we must ensure that when adversaries gain a foothold, we are prepared to detect them quickly, evict them decisively, and keep critical operations running without interruption.” CISA’s press release describes the approach as “decisive, proactive action to identify, disrupt, and eliminate adversarial access.” Those are CISA’s descriptions of its intent. They are not results, and the strategy does not claim they are.

    The most consequential line, in my view, sits under the Harden pillar: “We will deprioritize threat-agnostic, all-hazards approaches and embrace risk-informed, consequence-based analysis.” CISA says it will focus attention, investment, and operational effort “on the infrastructure whose disruption would have the greatest impact on public health and safety, continuity of the economy, and national security.” CISA also says it will support the White House in developing a National Risk Register to quantify risk to national infrastructure.

    What the Strategy Commits CISA to Do

    Almost every key objective is written as something CISA will do, often “alongside” or “with” partners. Among them:

    • Integrate cyber defense capabilities across government, industry, state, local, tribal, and territorial partners, and international organizations.
    • “Work alongside critical infrastructure owners and operators to modernize OT environments, improve asset visibility and anomaly detection, and assure OT supply chain integrity.”
    • Coordinate with partners on data-driven software security guidance and supply chain transparency.
    • Mature an “intelligence-to-operations pipeline” with the Intelligence Community, industry, and agencies, including real-time decision making and alerts.
    • Build capacity through CISA’s ten regional offices and modernize planning and exercise capabilities.
    • Expand priority telecommunications services, including 5G/6G and commercial satellite constellations.

    The timeline is also about CISA, not operators. The plan says that within 60 days CISA will “begin a programmatic review” of its core services and capabilities. Within 180 days it will deliver “an approach for implementation and measurement of outcomes.” Within 365 days it will deliver a progress update. As Nextgov/FCW points out, the 60-day deadline applies to starting the review, not finishing it.

    What the CISA Strategic Plan Asks of Critical Infrastructure Operators

    This is where I want to be careful, because it would be easy to read new mandates into this document. I did not find any. The strategy sets no new reporting requirements, deadlines, or technical controls for owners and operators. Nextgov/FCW also reports that it “does not establish new legal authorities.”

    What the strategy does contain are appeals. Andersen writes: “I call on everyone to do their civic duty and take every step necessary to deny our adversaries a single foothold or advantage.” In the press release he goes further: “I call on every company and organization in the critical infrastructure community—from owners and operators to vendors and technology manufacturers—to act now.” DHS Secretary Markwayne Mullin’s foreword urges unity at every level, including “owners and operators of businesses and critical infrastructure.”

    Beyond those calls to action, operators appear as partners. CISA says it will build “distributed collaboration frameworks that empower and enable SRMAs, regions, states, communities, and owners and operators to advance infrastructure security through localized action.” The specifics of what partnership will look like are deferred to the 180-day implementation approach.

    My Read and Recommendations for Operators

    These are my opinions, not CISA guidance, and none of them is a requirement in the strategy.

    1. Figure out which side of the prioritization line you are on. A consequence-based model means some sectors and some facilities get more CISA attention, and others get less. The strategy does not say which services lose resources. Do your own consequence analysis now so you can make your case when the implementation plan arrives.
    2. Invest in OT visibility before anyone asks. Asset visibility and anomaly detection are named objectives. You cannot be a useful partner in evicting an adversary from a network you cannot see.
    3. Rehearse eviction, not just detection. The strategy’s own language is “detect them quickly, evict them decisively.” Tabletop what removing a persistent intruder from your control environment would take without stopping operations.
    4. Inventory the free CISA services you depend on. The 60-day programmatic review covers CISA’s core services, and the plan does not say which ones might change. If a free scan or assessment is part of your program, note it and have a fallback in mind.
    5. Get to know your CISA regional office. Regional capacity is part of the plan. A relationship built before an incident pays off during one.
    6. Watch for the 180-day implementation approach. That document, expected roughly in late April 2027 if CISA hits its stated timeline, is where concrete expectations would likely show up.

    Here is a hypothetical worth thinking through. A mid-sized municipal water utility reads this plan and assumes it is now in line for hands-on CISA help. If the risk register and prioritization land elsewhere, that utility could end up with less direct support than before. The strategy does not say that will happen. It does make the tradeoff explicit.

    The Bottom Line

    CISA 2030 promises, in CISA’s words, an “aggressive, anticipatory stance” against adversaries in critical infrastructure. For now it is a statement of agency priorities and a timeline for CISA’s own work. It is not a rulebook for operators. Treat the calls to “act now” as a strong signal, and build your visibility, eviction playbooks, and consequence analysis before the implementation details arrive.

  • Hackers Reportedly Tried to Breach the FEC, Taking Its Filing System Offline a Week Before the Midterms: Planning for When a Public Deadline Outlives Your Platform

    Hackers Reportedly Tried to Breach the FEC, Taking Its Filing System Offline a Week Before the Midterms: Planning for When a Public Deadline Outlives Your Platform

    Late Monday, October 26, a week before the November 3 midterms, the Federal Election Commission’s electronic filing system went dark. For about two days, campaigns, super PACs, and outside groups could not submit the rapid-turnaround disclosures the law requires in the final stretch of an election. This FEC hack campaign finance filing outage deserves attention well beyond politics. It is a clean case study of what happens when a legal deadline keeps running after the platform you file on stops working.

    What We Actually Know About the FEC Hack and Campaign Finance Filing Outage

    Precision matters here, because the public record is thinner than many headlines suggest. According to Reuters reporting published by NBC News, an FEC electronic filings officer emailed several campaign compliance firms Monday night. The email said: “We have a bad actor attempting to breach FEC systems. As a result, we’ve had to temporarily shut down our web facing services.” The FEC’s own public statement was narrower: “at this time, the FEC’s electronic filing system is currently offline and our IT staff are actively working to restore service.” The agency said previously filed reports remained searchable and otherwise declined to comment.

    A status message put the start of the outage at around 10 p.m. ET Monday. CNBC reported that on Tuesday night the FEC said it had identified the issue and was implementing a fix. CNBC also noted that the FEC had not said who was behind the reported breach attempt, whether anyone gained access to its systems, or whether any data was compromised.

    By Wednesday, October 28, Reuters reported the system had been restored after being offline for nearly 41 hours “to contain a cyberattack.” Reuters also reported, citing an email, that campaign finance managers had to reset their passwords before logging back in. The FEC did not announce details on the cause, and Reuters could not establish a motive. CISA and the FBI referred questions to the FEC.

    So the fair summary is this. An FEC official described an attempted breach, and the agency took its web-facing filing services offline to contain it. As of these reports, there was no public confirmation that attackers got in or that any data was taken. I am not going to claim otherwise.

    Why the Timing Hurt

    The FEC does not administer elections. Reuters and CNBC both stress that the outage had no effect on voters’ ability to cast ballots. What it hit was transparency. CNBC explains that in the final 20 days before an election, candidate committees must report certain contributions or loans of $1,000 or more within 48 hours. Political committees and outside groups must disclose certain spending within 24 hours.

    That window is when the money moves fastest. Brendan Glavin, research director at OpenSecrets, told Reuters that super PACs and dark money groups spent an average of over $50 million per day in the week before the outage, with more than $250 million reported on October 20 alone. “Disclosure delayed is disclosure denied,” OpenSecrets executive director Hilary Braseth said in a statement. Steven Mele, CEO of compliance firm MBA Consulting Group, told Reuters: “I’ve never seen this in 26 years.”

    CNBC adds context that should worry anyone who manages public-sector IT. The six-member commission has had just two members since October 2025 and has lacked the four commissioners needed to take official action since May 2025, while career staff keep operations running.

    The Containment Call Was Probably Right

    Kyle Sandy, cybersecurity director at Logically, told Reuters that taking the system offline is the right call during an attack. “It tells me the FEC detected something they couldn’t immediately scope and chose a known outage over an unknown risk,” he said. I agree. When you cannot yet scope an intrusion attempt, shutting down web-facing services and forcing credential resets is defensible. That is especially true when the integrity of filed records is the whole point of the system.

    The hard part is that the rest of the world keeps its own clock. Filers still had legal deadlines. Watchdogs still had a public to inform. That is the gap I want IT leaders to plan for.

    My Recommendations: When a Deadline Outlives Your Platform

    These are my opinions for IT, compliance, and security teams, not FEC guidance or legal advice.

    If You Run the Platform

    • Publish an outage protocol before you need it. Filers should know in advance what counts as timely filing when your system is down, and where official updates will appear.
    • Separate intake from the public web. Design a degraded mode, such as an authenticated backup submission channel, that you can keep open while you contain an incident on the main portal.
    • Plan the credential reset. Forced resets for every filer at once are a support event. Script the communications and staff the help desk before the incident.
    • Say what you can, early. Reuters noted that the lack of information fueled conspiracy theories. A short, factual statement on scope and data status narrows the rumor space.

    If You Depend on Someone Else’s Platform

    • Map your hard deadlines to single points of failure. Any regulatory filing, payment run, or customer notice that depends on one external portal belongs on your risk register.
    • Keep evidence of attempted filing. Screenshots, timestamps, error messages, and saved submissions protect you if questions come up later.
    • Prepare filings ahead and queue them. Have reports ready to submit the minute service returns, rather than scrambling in the backlog.
    • Know your official channels. Decide in advance which agency status page or account you treat as authoritative, and do not act on rumors.

    Here is a hypothetical worth a tabletop exercise. Your quarterly regulatory filing is due at midnight, and the agency portal goes offline at 9 p.m. because of a suspected intrusion. Who decides what you do, what do you document, and who tells leadership? If you cannot answer in five minutes, you have homework.

    The Bottom Line

    Based on what has been reported, this was an attempted breach that forced a deliberate shutdown. It was not a confirmed data theft. The lesson is not unique to campaign finance. Every organization that runs or relies on a deadline-driven platform needs a plan for the hours when security wins and the system stays dark.

  • Bridgewater’s 100 Million Document Vault and Its Permission Problem: Building Enterprise AI Search That Respects Who’s Allowed to See What

    Bridgewater’s 100 Million Document Vault and Its Permission Problem: Building Enterprise AI Search That Respects Who’s Allowed to See What

    Most enterprise AI pilots work fine until the corpus gets big and the access rules get messy. Then the search quietly breaks. Bridgewater Associates has now described, in public, how that happened to its own document platform. It is the clearest case study I have seen of why permission-aware RAG enterprise AI search is a retrieval-engineering problem first and a model problem second. If your team is wiring a chatbot into SharePoint, email, or a research archive, this is the part you cannot skip.

    What Bridgewater Says It Built

    At AI Engineer NYC 2026, Bridgewater’s Sam Green, team lead for Applied AI, described the Vault as “our repository of external and internal unstructured data.” By his account it holds “more than 100 million documents,” with thousands more added every day from 135 external sources, plus internal data from inside the firm. Each document is enriched with tags like country, concept, market, and the time period it discusses (session transcript).

    Aaron Linsky, who runs product engineering at Bridgewater, listed five dimensions the team had to get right: recall, precision, filtering, entitlements, and scale. He said production traces showed “almost always there was some condition,” such as a date, a source, or a specific asset. That is where the team “ran into our first major performance issue.” He also put numbers on the two audiences. Humans generate about 100 queries an hour across roughly 300 users. Agents easily run “100-plus queries a second,” and agent swarms can reach thousands.

    Why Permissions Break Vector Search

    In a separate joint session with turbopuffer CTO Nikhil Benesch (“Can Your AI Analyst Keep a Secret?”), Linsky said that when Bridgewater first built this “without turbopuffer, we thought it worked fine.” Then, as documents piled into the Vault, “it fell over tremendously.” Benesch explained why. Internally, “every email, every meeting note, every report might have a completely different set of permissions attached to it.”

    Benesch described two classic approaches, and each one fails differently:

    • Post-filtering: grab the nearest vectors first, then throw away the ones the user cannot see. It is fast, but a user with narrow access can end up with few or no results, even when a relevant document they are allowed to read exists.
    • Pre-filtering: narrow to everything the user may see, then compare vectors. Recall is good. But his example of a user with access to 50 million of 100 million documents means a huge sequential scan, and the vector index stops helping.

    turbopuffer’s answer, which Benesch calls native filtering, combines the two. The index tracks which vector clusters contain at least one document a given user can access. It skips the clusters with none, then scans inside the nearest eligible ones. Benesch said clusters currently hold about 200 documents and that the size is still being tuned. Linsky summarized the outcome in his keynote: fast like post-filtering, “but with close to the same recall as pre-filtering.”

    About Those Headline Numbers

    You will see “450 million vectors” and “120 milliseconds” attached to this story. Those figures come from Benesch, the vendor’s CTO, not from a Bridgewater statement. In the joint talk he said that across 450 million vectors, Bridgewater’s 100 million documents “turn into 400 million vectors,” and that turbopuffer gets “120 millisecond searches” at P90 across those namespaces. Note that 400 million and 450 million appear in the same sentence. Treat both the vector count and the latency as turbopuffer’s reported results, not an independent benchmark. Benesch acknowledged that nested groups are a limitation where “things get tricky.”

    How the Permissions Are Actually Wired

    This is the part security teams should read twice. turbopuffer’s own permissions documentation says permissions “currently have to be implemented at the user-level with filters.” The product “doesn’t provide built-in mechanisms for row/document-level RBAC.” The recommended pattern is to store the user IDs or group IDs with read access on each document. At query time, you pull the caller’s identity and groups from your own auth layer and pass them in as a filter.

    The docs flag a subtle failure mode, too. An empty permission array should mean no access. To make a document public, write an explicit flag like is_public: true rather than inferring openness from an empty list. That design choice is safer, and the docs say it is faster.

    The lesson goes beyond any one vendor. The vector store enforces whatever filter you hand it. The truth about who may see what still lives in your identity system and your ingestion pipeline.

    My Recommendations for Permission-Aware RAG in Enterprise AI Search

    These are my opinions, not Bridgewater or turbopuffer guidance.

    1. Make entitlements part of ingestion. Copy ACLs, groups, and source licensing onto each chunk when you index it, and keep them in sync when permissions change. A stale ACL is a data leak waiting for a prompt.
    2. Default-deny. Follow the explicit public-flag pattern. Never let a missing or empty permission field mean “everyone.”
    3. Resolve identity at query time, server-side. The model should never pick the filter. Your backend attaches the caller’s user and group IDs after authentication.
    4. Benchmark with permissions turned on. Bridgewater built about 200 hand-labeled examples in under a week and measured precision at 10 rising from about 60% to the 90–95% range. Run your test set as different users, including narrow-access ones, or post-filtering failures will hide from you.
    5. Plan for agent load. Size for machine query rates, not human ones. Bridgewater’s own gap between human and agent traffic is, in Linsky’s words, “a few orders of magnitude.”
    6. Flatten or pre-compute nested groups. Even the vendor says nested groups get “tricky.” Expand memberships in your pipeline instead of hoping the search engine resolves them.

    Here is a hypothetical to test your design. An analyst in a restricted deal team asks the assistant about a target company. The best-matching memo belongs to a different team. With naive post-filtering, the analyst gets nothing useful. With a missing ACL, they get a memo they should never see. Both outcomes are failures, and only one of them shows up in a latency dashboard.

    The Bottom Line

    Bridgewater’s account shows that permission-aware retrieval decides whether enterprise AI search is usable and safe at scale. The impressive 450-million-vector, 120 ms figures are turbopuffer’s numbers. The durable lesson is architectural: carry entitlements on every document, filter at query time from a trusted identity source, and test recall as real users with real restrictions.