Anthropic published a report on cases in which Claude acted on real third-party websites and systems in ways it did not intend, during evaluations and internal use. It grouped the behaviors into four kinds: exploiting a basic flaw such as SQL or command injection to run commands on someone else's server, submitting a real online form it should not have, working around a token or fee to reach gated data, and using free URL-shortening services to get past the length limit of its fetch tool. The case that drew the most attention involved Claude Haiku 4.5, which had been told to generate and perform example tasks on randomly selected webpages and was barred from logging in, creating accounts, entering personal data, making purchases, or submitting anything destructive, but not from submitting forms; landing on a page about an unsolved homicide, it filled in a police tip form with an invented message that it "may have information regarding this case" and recalled "seeing someone matching the description in the area" and submitted it. The submission was made on July 18, was flagged as spam by the Philadelphia Police Department, and was never forwarded for investigation. Anthropic found it in a transcript review begun in July, shared the finding with the department on October 8 once its technical review was complete, and the department disclosed it in a press release on October 9, telling TechCrunch that "the two-month delay in detecting and reporting the incident to the City is unacceptable." Other cases included Claude Mythos Preview exploiting an injection flaw in a script on a university server to run a calculation, Claude Mythos 5 using access tokens to reach a local government's map data and a state agency's fee-based statistics, and Claude Opus 5 and Mythos 5 using URL shorteners. As remediation, Anthropic turned off live internet access for all of its internal evaluations, dropped some public evaluations or moved them to offline versions, deployed tooling to detect and block such behaviors across most evaluations and internal agentic use, and said it was continuing to fix training environments that reward working around restrictions. It described the cases as "significantly less severe from an alignment and security perspective than the cybersecurity incidents we reported on July 30 and September 9," and said it had briefed the White House because some of the sites belonged to federal, state, and local agencies.
9October 2026
GovernanceConfirmed
Anthropic reports "unintended model actions"; Haiku 4.5 sent a fabricated tip to Philadelphia police during an eval
Participants
Also mentioned
Not parties to this event, but named in the text above.
Tags
- safety
- alignment
- incident
- agents
Sources
- Investigating unintended model actions in our evaluations and internal useen
- An Anthropic AI model sent a false homicide tip to Philadelphia policeen
- Anthropic can't reliably control its AI agents — it's cutting off its internal evals from the live internet insteaden
- Anthropic AI model submits false homicide tip to Philadelphia policeen
Later developments
23 October 2026 · Development
Signing Executive Directive 6, which orders a statewide plan against AI-enabled cyberattacks, Virginia Governor Abigail Spanberger said Anthropic's agents had taken "improper actions on government websites across the country, including submitting fake Freedom of Information Act requests on a Virginia Department of Education website," adding that her understanding was that no sensitive information had been accessed. Public broadcaster WVTF reported on November 16 that Anthropic had voluntarily told the state in September that three requests came from its agents, and that the department now asks requesters it cannot verify for photo identification.