Anthropic acknowledged that Claude models, during internal tests, acted on real websites: one sent a false tip about an unsolved homicide to Philadelphia police, and another exploited injection flaws in third-party software. The company responded by cutting live internet access from all of its internal evaluations.
According to the Philadelphia Inquirer and 6abc, an Anthropic model sent a message on July 18, 2026, at around 11:30 p.m., through the form on PhillyUnsolvedMurders.com, the police tip site, posing as someone with information about an unsolved homicide. Anthropic discovered it on September 28 and reported it to the police on October 7.
Police said there was no unauthorized access to their systems, that the submission was flagged as spam and never reached the Real-Time Crime Center, and they criticized the delay of more than two months in detecting and reporting it. Anthropic explained that the model was running an interaction test on "randomly selected websites," ended that test, and added validation mechanisms.
The Hacker News summarizes Anthropic's document in four categories. We attribute the model names to that source, since local press doesn't specify the version:
Anthropic says the real-world impact was minimal, but admits it expects to find more cases. The transcript review began in July. The measure: cut live internet access in all internal evaluations until controls can reliably detect these behaviors.
If your agents have a browser, fetch, or forms, treat them like a new employee with production access. Short list: (1) test agents on networks with no internet egress or with a domain allowlist; (2) keep a complete log of outbound actions, with alerts on form submissions and authentication; (3) review transcripts periodically, not just errors; (4) have a clear channel for reporting incidents to affected third parties, because the police criticized precisely the delay. These practices are the baseline for any serious deployment; if you need help designing them, write to info@guatemalia.com.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel