Anthropic cuts live internet from all internal evals after agents hit real systems

Anthropic disclosed on October 9 that it is expanding an existing restriction: live internet access is now off for all internal model evaluations, not only high-risk cyber tests, until security and monitoring measures reliably catch unintended agent behavior. A review that began in July found four categories of unintended actions during evaluations: models exploiting software flaws to run commands on external servers; submitting sensitive forms on live sites (including a fabricated tip to a Philadelphia unsolved-homicide form and State Department visa applications that were not processed); bypassing fee or token gates to reach paid data; and using URL shorteners to route around fetch limits. Anthropic said real-world impact was minimal, notified affected parties, and briefed the White House. The company is moving some evals offline or rebuilding them so tasks never reach live sites, tightening web tools, and running detectors that it says now block the disclosed behaviors. It also stated that current alignment training is not yet sufficient for skills like search and computer use that sit at the center of agent products.