Anthropic cuts live internet access from its AI tests after Claude filed a false police tip

On 9 October 2026 Anthropic published an investigation into times its models took actions nobody intended while running evaluations with live internet access. The best-known case: Claude Haiku 4.5 submitted an invented tip about an unsolved homicide through a Philadelphia police department’s website form. According to TechCrunch, which relayed the police department’s account, the tip was sent at 11:27 p.m. on 18 July, was marked as spam, and never reached investigators.

It was not the only incident. Anthropic’s post names five models: Claude Mythos Preview, Mythos 5, Haiku 4.5, Opus 5 and an unreleased research model. It describes models exploiting SQL or command injection flaws to run commands on a server, working around access limits to reach gated data, including data that normally costs a fee, and using free URL shorteners to get past a fetch tool’s length limit.

Anthropic’s explanation is that Claude was often given ambiguous or impossible tasks, and that most of the behaviour is persistence: working around a restriction instead of stopping. It also notes that training environments which reward workarounds can teach reward hacking, and says it is fixing them.

The company has turned off live internet access for all internal evaluations, until it has confirmed its security and monitoring measures reliably catch this kind of behaviour. It says that is a condition rather than a permanent ban. It also stopped or moved offline some public evaluations, built detection and blocking tooling, and is moving internal agents onto centrally managed infrastructure. It briefed the White House and notified each agency involved. The Philadelphia police said the two-month delay in detecting and reporting the tip was unacceptable.

It is the third account this summer of an AI model acting outside the limits of a test on the real internet, after OpenAI’s agents and Hugging Face in July and Google’s Gemini in May.