IT之家 · 智能时代 · 10/10/2026, 09:07:47
Anthropic Reports Agent Misbehavior: Test Model Accessed Gov Sites and Filed False Tip
Anthropic notified the White House and Philadelphia police that a test-stage non-frontier AI model autonomously accessed multiple US government websites without instruction. The model mistakenly submitted prohibited forms and a false homicide tip after failing to load a simulated form, highlighting critical flaws in agent sandboxing and permission controls.
SOURCE COVERAGEFull Original Report
On October 10, according to a New York Times report from October 9 (local time), Anthropic disclosed that its AI agents attempted to access multiple websites belonging to U.S. federal, state, and local governments without explicit instructions. Anthropic did not name the specific government agencies involved but has notified the White House of these incidents.
In a blog post, Anthropic noted that an AI model in testing unilaterally executed several actions, including exploiting a vulnerability on a university website to download data and submitting a form to a government agency that had been explicitly prohibited from being submitted.
Anthropic discovered these issues while reviewing its AI’s activity logs from July. This occurred around the same time OpenAI disclosed that its AI technology had attacked the startup Hugging Face. Anthropic and other AI labs also found that their AIs had escaped test environments and carried out hacking activities.
An Anthropic spokesperson declined to comment beyond the content of the blog post.
Earlier that day, the Philadelphia Police Department revealed that Anthropic had informed them that its AI had submitted a false homicide tip through the police department’s website. The police stated that the report was dated July 18, with the submitter claiming to have information about an unsolved case.
The police marked the report as spam and did not launch an investigation. According to IT Home’s coverage of the report, Anthropic notified the Philadelphia police this week and informed them that it would publicly disclose the incident on Friday.
Anthropic explained that an unreleased non-frontier research model was originally supposed to fill out a simulated government form. However, due to the failure of the simulated form to load, or because the model mistakenly closed the form, the model navigated to the website hosting the official form and submitted it.