Sriram Krishnan, the former senior White House AI policy adviser, appeared on CNBC on July 1, 2026, to discuss Anthropic’s Claude Mythos and its ability to find and exploit software vulnerabilities.
His message was blunt. Mythos is effective at discovering cyber weaknesses, and critical systems need protection. The statement came the same day that US export controls on Anthropic’s Mythos 5 and the more guarded Fable 5 variant were lifted, barely weeks after they were imposed.
From internal leak to export controls and back again
The model was leaked internally in late March 2026, showcasing what observers described as a step-change in autonomous vulnerability discovery. That leak immediately triggered national security concerns and drew scrutiny from White House officials.
Anthropic officially launched Claude Mythos Preview on April 7, 2026. Through a limited initiative called Project Glasswing, the company granted early access to roughly 50 trusted partners, including banks and security firms. The results were staggering: Mythos uncovered over 10,000 critical software vulnerabilities through that restricted program alone.
