The White House Says AI Companies Must Report Model Incidents: What It Means for Your Own Agent Logs

On October 9, Axios ran a story headlined “Exclusive: Anthropic breaches spark White House AI reporting mandate.” Since then, “AI incident reporting mandate” has been the shorthand for what happened. The real picture is narrower and looser than that phrase suggests. It is still worth an IT team’s attention, mostly because of what it implies about the logs you keep on your own AI agents.

What the Statement Actually Says

The text comes from a statement by leaders of the White House Super Intelligence Force (SI Force), which Axios says was “shared exclusively with Axios.” Axios published it in full in its October 9 article. Every quote in this section comes from that statement as Axios published it.

The core lines:

• “SI companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm.”
• “This notification and remediation process is not optional. It is a critical national security obligation.”
• “We expect Anthropic and all companies to honor their obligations, immediately report incidents, fully cooperate with federal and state law enforcement authorities, remedy any damage, and implement concrete safeguards to ensure these failures do not reoccur.”
• “Our message to all SI companies is clear: delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated.”

(“SI” is the administration’s term for AI. A September executive order directs agencies to use “Super Intelligence” in place of “Artificial Intelligence.”)

As for what prompted it, the statement says: “Earlier today, Anthropic contacted the SI Force to disclose the details of various prior incidents that it discovered in late September involving the unauthorized and fraudulent use of government and other systems.” It adds that “The company informed us that these events occurred in the past, the activity has ceased, and there is no ongoing similar activity.” I am leaving the characterization there. The quoted words are the statement’s, not my own findings.

What It Does Not Say

This is the part that matters for anyone deciding whether to change process. As of October 13, 2026, I could not find any executive order, OMB memo, or Federal Register rule that puts this requirement in writing. I checked whitehouse.gov’s presidential actions and statements and searched the Federal Register. The only text is the statement Axios published.

Read closely, the statement also leaves out the pieces a real reporting regime would need:

• No definition of an incident. It refers only to “incidents involving their models.”
• No deadline beyond “immediately.” There is no hour count.
• No reporting channel. It does not say where, how, or to which office a report goes.
• No penalties. Axios notes that “The statement did not make clear what enforcement mechanisms or penalties would look like if AI companies failed to disclose incidents and remediate them.”

Scope matters too. Axios reports that “The White House requirements apply to all AI companies,” and the statement speaks to “SI companies” and their models. Nothing in it addresses organizations that simply use AI tools or run agents built on someone else’s model. If you run an IT shop, the statement is not addressed to you.

Background: The 72-Hour Clause That Came and Went

Federal AI incident reporting has been tried before, in a narrower form. OMB memo M-24-18 (September 24, 2024) told agencies to, “to the greatest extent practicable, include contractual terms requiring that vendors have a process for identifying and disclosing to agencies serious AI incidents and malfunctions of the acquired AI system, or service within 72 hours, or a timely manner based on the severity of the incident.” It left the definition to agencies: “Criteria for what constitutes a serious AI incident or malfunction should be determined by each agency.”

That was a contract clause for federal purchases, not a rule for industry. And it did not last. OMB memo M-25-22, dated April 3, 2025, states: “This memorandum rescinds and replaces OMB Memorandum M-24-18.” I searched M-25-22’s full text and did not find a comparable incident-reporting clause.

Why an AI Incident Reporting Mandate Still Points at Your Agent Logs

Everything from here down is my own recommendation. The statement does not require any of it, and it does not apply to companies that only use AI.

Here is why I would act anyway. The message from Washington is that fast disclosure of AI incidents is now expected, at least from model makers. Expectations like that tend to flow downstream through contracts, vendor questionnaires and cyber insurance forms. When that happens, the first question will be simple: what did your agent do, and can you prove it?

Picture this hypothetical. An internal agent with web access submits a form to an outside site it was never meant to touch. Your vendor calls and asks for a timeline. If your logs show only “agent ran,” you cannot answer. If they show every action, with inputs and approvals, you can answer in an afternoon.

What I would log for every AI agent you run:

1. Identity. Which agent, which model and version, and which human or service account it acted for.
2. Every external action. Outbound requests, form submissions, emails sent, files written and API calls, with timestamps and destinations.
3. Tool calls and permissions. Which tools it invoked, with what arguments, and whether a human approved them.
4. Prompts and key outputs. Enough to reconstruct why it did something, stored with the same care as other sensitive logs.
5. Blocks and failures. Refused actions and guardrail hits are often your earliest warning.

Then put the logs somewhere useful. Send them to the same SIEM as everything else, retain them long enough to reconstruct a timeline weeks later, and lock them so the agent itself cannot edit or delete them.

Finally, write the playbook before you need it. Decide now what counts as an agent incident in your shop, who gets told, and how fast. Neither the statement nor any rule defines that for you, so your own definition is the one that will matter.

The Bottom Line

What exists today is a strongly worded White House expectation aimed at AI companies, published through one outlet, with no written rule behind it yet. That could change. Either way, being able to say exactly what your agents did is good security practice on its own terms. That is my view, and it is the reason I would start logging now.