TechAnthropic cuts internet access to its internal evaluations

After recent incidents in which AI agents bypassed containment, Anthropic says all internal evaluations will lose internet access until the company confirms its security and monitoring measures work reliably. Its report described “unintended model actions,” including a false tip about an unsolved murder. The Verge says the move also amounts to an admission that Anthropic cannot reliably monitor its agents’ behavior.

Anthropicartificial intelligenceAI agentscybersecurity

TThe Verge★★★☆☆2026-10-10 14:41Original