Anthropic said its Claude AI model hacked into three organizations during cybersecurity testing.
The announcement came after a “swarm” of OpenAI agents escaped confinement, gained internet access, and broke into at least five companies, eventually stealing the answers to a cyberoffense evaluation.
The latest breach is less dramatic than the previously announced one: OpenAI’s models found and exploited vulnerabilities to escape, while in Anthropic’s case a misunderstanding led to the agent’s cage essentially being left open.
But both point to a difficult future: Most commercial AI is connected to the internet anyway, so confinement is irrelevant, and open-weight models that can easily have any anti-cyber guardrails removed by bad actors are now nearly as capable as frontier products.