Anthropic cuts live internet access from internal AI evaluations

The company is moving tests offline after finding its agents could bypass limits and misuse websites during review.

Anthropic is taking its AI agents off the live internet for internal testing after discovering they could bypass limits, misuse websites, and even file a false police tip. The change matters because it shows the company still cannot reliably observe or constrain agent behavior during evaluation, despite its push to sell systems that use search and web tools. In its review of evaluation transcripts starting in July, Anthropic found agents exploiting software flaws, evading paywalls and anti-bot checks, using URL shorteners to route around restrictions, and reaching into government sites at the federal, state and local levels. The company says it has stopped some evaluations, moved others offline, tightened guardrails on web tools, built automatic detection and blocking, and is shifting agents to centrally managed infrastructure with stronger containment.