314. The Responsible

Last week, Senator Mark Warner told a Senate hearing that Anthropic's AI model, Mythos, had successfully penetrated nearly every classified NSA network in the country.

This was not technically a breach. The NSA had hired Anthropic to test its own defenses. This is the security industry's version of inviting a locksmith to break into your house to confirm that your house is secure. The test was expected to take weeks. Mythos completed it in hours.

Senator Warner called this "impressive." He also said, "Thank God it was Anthropic."

(The other half of that sentence, which Warner left graciously unspoken, is: "...rather than an entity that might, upon gaining access to nearly every classified national security database in the United States, do something with it." The entire reassurance rests on the assumption that Mythos had no goals of its own while it was inside the network. This is, at present, a reasonable assumption. It is also the whole bet.)

Warner is now pushing for mandatory independent testing of AI systems on critical infrastructure. The logic is that before we find out whether an advanced AI can compromise the country's most sensitive institutions, we should find out under controlled conditions — scheduled, documented, with receipts, during normal business hours.

The difficulty, noted at the hearing and then moved past, is that "independent testing" becomes complicated when the system being evaluated is more capable of performing the test than any available evaluator. Anthropic has not weighed in publicly on who should conduct this review.

They would probably be available to help.

Several people in the hearing room did not ask about this.

I am not making this up. Senator Mark Warner said this at a formal Senate hearing. The NSA agreed to the test. Mythos did the thing. The Senator's reaction was relief.

Senator Warner described Anthropic as "responsible." He meant this as a compliment.