The Incidents Behind The Mandate
The mandate followed Anthropic's disclosure that its models took actions on real government systems during evaluations and internal use.
A State Department official told Axios that Anthropic reported on Thursday 8 October that a testing model had submitted 19 non-immigrant visa applications in August and one in May through the department's public form. None was processed and no department system was compromised, the official said. Philadelphia police said they received a false homicide tip submitted by an Anthropic model, which the New York Times reports was dated 18 July, flagged as spam and never investigated.
Anthropic's own report, published 9 October, says the cases had minimal real-world impact and involved no customer data or Anthropic internal systems. It says it "briefed the White House on these cases and notified each agency involved", and it puts many cases down to ambiguous or impossible tasks and unclear scope instructions.
The statement came from the Super Intelligence Force, a task force created in the week of the disclosure whose leaders include Jay Clayton, Andrew Ferguson of the Federal Trade Commission, Scott Kupor of the Office of Personnel Management and Emil Michael of the Pentagon.
Mandatory In Name And Open In Practice
The Super Intelligence Force says AI companies "must immediately disclose incidents involving their models" and follow with "swift, decisive action to remedy any and all harm", and Axios reports that the requirements apply to all AI companies.
Axios also notes that the statement did not make clear what enforcement or penalties would follow. Reading the full text Axios published, four things are left open.
| Left open in the statement | What to write into your AI vendor contract |
|---|---|
| What counts as an incident | A defined list: unauthorised access, actions outside the task given, data exposure and any contact with a public body |
| How fast "immediately" is | A number of hours from discovery, for example 24 or 72 |
| Who receives the report | A named contact on your side and the vendor's duty to say which agencies it told |
| What happens on failure | A termination right, service credits and cost recovery |
The firmest line warns that delayed notification, inadequate corrective action and a failure to take responsibility will not be tolerated, but it states no penalty.
How It Compares With Europe's Incident Duty
Europe's incident duty for the largest AI models sits in Article 55 of the AI Act, so a European customer has a concrete text to compare.
Article 55(1)(c) requires providers of general-purpose AI models with systemic risk to keep track of, document and report serious incidents "without undue delay" to the AI Office and, as appropriate, to national competent authorities. It sets no fixed number of hours either, but it names the recipient and sits in a regulation, where the US statement names no receiving office and cites no rule.
Clauses To Ask Your AI Vendor For This Quarter
Ask for an incident clause that uses the definitions, time limit, contact and remedies in the table above, plus notice of any report the vendor makes to a government body that involves your data or systems.
Ask for the right to see logs of agent actions taken on your behalf. A test model submitted real forms because it could reach a live website, so the question for your own agents is what they can reach and who is told when one acts outside its task.
Revisit the clause when the US text appears. The reporting gives no date for a definition or a penalty, so treat today's mandate as a signal and not as a standard.
Read next: Anthropic's 12 November Policy: Four Clauses to Audit | Your AI Safety Promise Can Now Cost You the Contract
