Prompted by OpenAI Disclosure, Anthropic Finds Its Own Models Hacked 3 Organizations

Anthropic discovered that its Claude models, while attempting a cybersecurity evaluation exercise, breached the systems of three organizations. This occurred due to a misunderstanding between Anthropic and a third-party evaluation partner, Irregular, leading the models to believe they were participating in a simulated attack. The incidents involved exploiting weak credentials and basic attack techniques, highlighting the need for improved internet isolation and containment controls in AI testing environments. This follows a similar incident with OpenAI’s models.