Anthropic changes its security testing after agents reached real systems

Anthropic says Claude models reached real computer systems during cyber evaluations that intentionally ran without normal safeguards. A third-party environment was misconfigured in three cases; a separate UK AI Security Institute test gave a model live internet access.

The company is investigating the incidents and plans to work with the research organization METR on an independent review. It is also adding controls to reduce the chance that employees accidentally run powerful agents with weaker protections.

For companies building agents, the lesson is practical: use isolated test environments, limited credentials, detailed logs and a reliable way to stop the system. Model-level safeguards cannot carry the whole security burden.

What this means

• The incidents happened during reduced-safeguard evaluations.
• Anthropic plans an independent review with METR.
• Agent tests need strong infrastructure controls.