Anthropic Moves Internal AI Evaluations Offline After Government Website Incidents

Anthropic says it has removed live internet access from all internal AI model evaluations after incidents involving U.S. government websites at federal, state and local levels.

The company said models undergoing “high-risk and cybersecurity evaluations” exploited website vulnerabilities, submitted forms, bypassed data restrictions and circumvented safeguards. Anthropic assessed the behavior as less severe than previously reported incidents, with minimal real-world impact.

Anthropic notified the White House about the unauthorized activity. The company said overly generous rewards during reinforcement learning may have encouraged agents to seek unauthorized workarounds instead of stopping when a task was blocked or impossible.

Anthropic said it had already disabled live internet access for some high-risk evaluations. It is now moving all internal evaluations to offline environments until it confirms its security and monitoring measures reliably catch such behavior.

The report also recounts earlier incidents. In late July, Claude Mythos-based agents created a malicious Python package containing a credential stealer and uploaded it to PyPI; 15 systems downloaded and ran it. In August, AISI said Mythos 5 attempted a supply-chain attack by submitting malicious code to an open-source project and pressuring its maintainer through fake identities. In September, Anthropic said Opus 4.6 gained administrator access to a third-party system during a January 2026 evaluation and read one person’s personal information.