Anthropic's Logan Graham corrects the Washington Post account of the Cathedral intrusion

Logan Graham of Anthropic's Frontier Red Team posted on October 26 corrections to the Washington Post story and the reaction to it on X, which said a security startup's AI agents got into Anthropic's systems in under 12 hours. These are his statements, not an Anthropic announcement:

• It did not take 12 hours: Anthropic saw several days of attempts. After them, the group exploited two third-party vulnerabilities; one it reported, and Anthropic found the other independently and cut off access.
• Nothing came close to Anthropic's most sensitive systems, such as model weights or Claude conversations.
• It was not an ordinary bug-bounty finding: it involved exploiting third parties and accessing data. He calls this unusual and outside the scope of Anthropic's safe-harbor policy, and says Anthropic is grateful the issues were found.

Graham also gives his own view that this is what the future of cybersecurity looks like: adversaries will use whatever models and tools they can, target everything with agents around the clock and hit dependencies, and he thinks models half to one generation behind the frontier can do this. He suggests red-teaming yourself continuously and redesigning systems to be less exposed, pointing to Anthropic's recent research on formal verification and attack-surface reduction.

The post had 48 points and 4 comments on X by October 26.