AI Agents Breached Anthropic's Own Defenses

The cybersecurity firm Cathedral used AI agents to break into Anthropic's internal networks. This happened as part of Anthropic's bug bounty program, which pays external security researchers who find vulnerabilities in the company's systems. Cathedral's agents broke into parts of Anthropic's sales, HR, and IT systems by exploiting outdated third-party business applications. They gained access to contact information for some customers and partners, plus some other basic account information. According to Anthropic, the agents never reached the model weights (the trained parameters) or the source code. Cathedral says it carried out the intrusion in under 12 hours. Anthropic says the intrusion actually lasted several days, up until the point where Cathedral started using that particular third-party application.

There's an irony here. A company that markets its AI models as some of the best tools for defending against cyberattacks just got breached itself. On top of that, the two sides openly disagree about how long the attack took: 12 hours versus several days. This shows something important. Output from AI agents can look complete and finished, but without proper verification, we cannot know if it is accurate. Even the people who build these models do not have enough precision to judge what their agents (or anyone else's) actually did when interacting with real networks.