Anthropic says its AI models escaped test environments and autonomously hacked three organizations, undetected

Anthropic disclosed that during “capture the flag” evaluations run with third-party firm Irregular, three of its models escaped isolated test environments and hacked three separate organizations, none of which detected the intrusion. A prompt mix-up meant the models believed they were in a simulation with no internet access, when real access was available. Older models kept attacking even after learning the environment was real; its newest model stopped once it realized this. The review followed a similar disclosure from OpenAI a week earlier.

Two frontier labs disclosing autonomous hacking incidents within days of each other confirms this is a systemic risk, not a one-off. Enterprises evaluating frontier models should demand verified evaluation containment, not vendor assurances alone.