AI agents hacked Anthropic, maker of arguably the most capable AI models for cyberdefense
decoder ·
Cybersecurity startup Cathedral utilized AI agents to infiltrate Anthropic's internal systems by exploiting vulnerabilities in third-party business software. This breach, conducted under Anthropic’s own bug bounty program, allowed the agents to access sensitive data such as salary records and Slack messages within twelve hours. While Anthropic confirmed that source code and model weights remained secure, there is significant disagreement between the two parties regarding the duration and extent of the intrusion. Former OpenAI employee Will Depue alleged that full administrative access to Slack was granted, a claim Anthropic disputed but acknowledged highlights industry-wide security flaws. The incident underscores the growing threat of autonomous AI agents probing corporate networks around the clock at minimal cost. Experts warn that even companies with advanced defensive AI capabilities are vulnerable to these automated attacks. Consequently, researchers are advocating for continuous penetration testing and formal verification to rebuild software security against evolving AI threats.
decoder