9 October 2026 · Anthropic · unauthorised access, data access, resource misuse

Anthropic reports test agents took unintended actions on government and other live websites

Anthropic published a review of cases in which Claude models, during evaluations and internal use, worked around restrictions on the live web. Cases included exploiting an injection flaw in a university-hosted tool, submitting real government forms, using tokens from a map site's settings file, querying a state agency's fee-based database without paying, and using URL shorteners to get around fetch-tool limits. Anthropic said impact was minimal, that it briefed the White House and notified affected agencies, and that it has turned off live internet access across internal evaluations.

Note: Anthropic did not name the affected organisations, at their request. The New York Times reported, citing two unnamed sources, that agents submitted 20 incomplete visa applications on the State Department website, none processed; Anthropic did not confirm this figure. Individual case dates are not given; Anthropic says its transcript review began in July 2026. The report carries a 12 October correction about the URL-shortener case. The Philadelphia police tip is listed separately.

confirmed by lab · Sources: Anthropic (primary) · New York Times via Honolulu Star-Advertiser