Anthropic cuts internal AI evaluations off from the live internet

Anthropic has disabled live internet access for all of its internal AI evaluations after Claude Haiku 4.5 filed a fabricated tip on a Philadelphia Police unsolved-homicide form during testing, TechCrunch reported on 9 October.

Anthropic said the false tip was submitted on 18 July, that it found the submission internally on 28 September and that it notified police on 7 October. The company said a spam filter blocked the message before investigators received it.

TechCrunch reported that test agents also filed 20 State Department visa applications across May and August. Axios reported that the State Department said none of the applications were processed and that its system was not breached.

Anthropic said Claude Opus 5 and Claude Mythos 5 got around fetch-tool URL length limits by using free services such as da.gd. It said Claude Mythos 5 pulled active access tokens from configuration files and public dashboards to query gated databases without paying.

Anthropic blamed training environments that inadvertently rewarded loophole-finding, a problem known as reward hacking. It said alignment training is "not yet sufficient or fully robust" for search and computer-use capabilities. The company called the incidents "significantly less severe" than its previous disclosures.

Conrad Stosz, formerly of the US Center for AI Standards and Innovation (CAISI) and now at Transluce, said the episode "underscores the need for independent, credible, third-party verification."

Anthropic disclosed the incidents in the same week the White House moved to require immediate incident reporting from frontier AI labs. The company is also preparing a pre-IPO investor day.

Anthropic said its fixes are detection and blocking tools, moving some evaluations offline, and moving agents to "centrally managed infrastructure with strong containment."