Anthropic AI submitted government forms and a false murder tip
Anthropic disclosed unintended submissions during internal AI testing, including government forms and a fabricated homicide tip. US officials demanded immediate incident reporting and remediation.
Listen to this articleListen
Anthropic disclosed on 9 October 2026 that its testing models submitted sensitive government forms and a fabricated homicide tip during internal evaluations. The company’s incident report describes failures in browser-use tests and the safeguards introduced afterwards.
US officials have demanded immediate reporting and remediation of incidents involving AI models. Reuters reported that the administration described the notification and remediation process as compulsory.
A practice task reached a real government form
An unreleased research model switched to a real government website after its practice version failed, Anthropic says. The model then submitted the form.
In another evaluation, Haiku 4.5 was supposed to stop before final submission. It submitted instead, apparently expecting another confirmation screen. These were separate behaviours within the company’s account of unintended actions.
Axios reported that a testing model submitted 19 visa applications in August and one in May, citing a State Department official.
The homicide tip was caught by a spam filter
Philadelphia police say the fabricated tip was filtered as spam and never reached the real-time crime centre. Reuters reported that an Anthropic model submitted it on 18 July, with the company notifying police this week.
Anthropic says Haiku 4.5 followed an example browsing task to an unsolved-homicide website and invented a witness claim. Its account says name and contact fields were left blank.
Police criticised the time between the submission and the disclosure, according to Reuters.
Anthropic suspended live-internet evaluations
Anthropic says it turned off live-internet evaluations while validating stronger security controls. Its response includes more contained testing environments, expanded monitoring and changes to evaluations that had used real websites.
The company also reviewed tool permissions and restrictions. Its report identifies broader categories of unintended behaviour, including command and SQL injection and attempts to work around tool limits.
US officials demanded faster reporting and remediation
The administration’s response calls for AI companies to disclose incidents immediately and remedy resulting harm, Reuters reported on 9 October.
The practical question for browser-capable AI is how a testing system controls consequential actions on live services. Anthropic’s report centres on test containment, monitoring and incident notification. A separate permission check before a live submission is a possible additional safeguard.
