A viral claim that Anthropic's Mythos AI model breached almost all NSA classified systems has been rejected by BitGo CEO Mike Belshe and others as misleading. The story originated from a Senator Mark Warner quote published in The Economist, where he recounted NSA Director General Joshua Rudd describing the tool's capability. However, missing context reveals this was an authorized red-team test on the NSA's own networks, not an external intrusion. The Economist editor later cautioned against literal interpretation, noting it depended on specific conditions and tool combinations. The US government was already a partner with Anthropic through Project Glasswing since April. Belshe bluntly called BS on the claim, joined by analysts who noted the breach was a test and that a separate jailbreak flagged by Amazon was the real concern. Anthropic stated the jailbreak only asked the model to find bugs, revealing minor known issues. The model's removal on June 12 was due to export controls, not security breach. The incident underscores the need for accurate reporting on AI security tests.

Key facts

• BitGo CEO Mike Belshe rejects viral claim that Mythos breached NSA systems as false.
• Claim originated from Senator Warner's quote, but was an authorized red-team test.
• NSA already partnered with Anthropic via Project Glasswing since April.
• Model removal on June 12 due to export controls, not security breach.
• Anthropic says it is working to restore access and drafting risk framework with White House.