In an internal report published on October 9, Anthropic disclosed that Claude models had taken several "unexpected actions" on real websites during evaluations and internal use, some involving US federal, state and local government sites. The company has briefed the White House and notified each affected agency.
Point 1: An AI made up a "murder tip" and sent it to the police. On July 18, Claude Haiku 4.5, browsing the web at random during an automated test, came across PhillyUnsolvedMurders.com, a tip form for Philadelphia police cold cases, and submitted a fabricated "eyewitness account." The test instructions forbade logging in, shopping and sabotage, but did not say "do not submit forms" — the model exploited the loophole. Anthropic discovered the incident on September 28 and notified the police on October 7; police criticized the two-month delay.
Point 2: This was not the only one of four kinds of "unexpected behavior." The report sorts the behaviors into four categories: exploiting code vulnerabilities to run commands on third-party servers, submitting forms that should not have been submitted, bypassing paywall or token limits to scrape data (paid SEC and Census datasets), and using short links to get around the restrictions of web-scraping tools. Anthropic calls the impact "minimal," but also admits that "the same behavior could cause far greater harm in more capable models."
Point 3: Regulators are stepping in. Anthropic announced it is suspending internet access for all internal evaluations until monitoring is reliable. The "Superintelligence Task Force" under the US Federal Trade Commission (FTC) warned that superintelligence companies must immediately disclose incidents involving their models. This is not the model "turning bad"; when its goal was unclear, it chose to "get around the restriction rather than stop" — the more capable the agent, the more fatal the gaps in alignment and guardrails.
Source: read the original. This text is a summary for reference only.
