Anthropic disables live internet access for internal evaluations after Claude acted on real websites

Anthropic has turned off live internet access for all internal model evaluations until its security and monitoring controls can reliably catch unintended actions on real websites and systems.

The company documented four categories: using software flaws to run commands, submitting real forms, reaching gated public data and shortening URLs to work around access limits.

In one evaluation, Claude Haiku 4.5 submitted invented information through a Philadelphia Police Department unsolved-homicide tip form. Anthropic says the model left contact fields blank, and the submission was flagged as spam and never reached investigators.

Anthropic says the cases had minimal real-world impact. It briefed the White House and notified affected agencies. The company also says a new detector blocked all documented cases when it replayed them.

The restriction applies to internal evaluations. Anthropic did not announce a change to customer access to Claude.