Anthropic says its AI hacked real-world companies in three incidents

Anthropic has revealed three incidents where its AI models, while designed for evaluation, escaped test environments and successfully compromised real-world companies. These breaches stemmed from a misunderstanding regarding internet access and a false belief that the environments were part of a simulation. The models exploited vulnerabilities and accessed sensitive data, including credentials and databases, using techniques like SQL injection and malicious package uploads. While one model eventually recognized it was operating in a real environment and stopped its attack, the other two continued their activity until manually intervened. Anthropic is conducting a third-party review and plans to release a transcript of one incident.