Anthropic's unreleased frontier model "Claude Mythos" reportedly breached nearly all of the U.S. National Security Agency's (NSA) classified systems within hours during an NSA-authorized red-team evaluation on June 11, 2026. This was not an unauthorized attack but part of a vulnerability test the NSA itself conducted.
June 14, 2026 · Anthropic · Claude Mythos
An AI broke into nearly all of the NSA's classified systems — in hours
In an authorized red-team exercise run by the NSA itself , Anthropic's cybersecurity model "Claude Mythos" infiltrated almost every classified system of the NSA and U.S. Cyber Command within hours — collapsing the path from vulnerability to weaponization from years to hours.
~All
Classified NSA / Cyber Command systems infiltrated
Hours
Time to breach in the authorized exercise
10,000+
Vulnerabilities found across Project Glasswing partners
17 yrs
Age of a FreeBSD RCE bug found & exploited from one prompt
Vulnerability → weaponization: the timeline collapses
The industry's warning: AI is compressing the path from discovery to attack from a scale of years down to a scale of hours.
How it unfolded
NSA uses Mythos for offensive cyber ops — despite DOD flagging Anthropic a supply-chain risk
→
NSA runs an authorized red-team exercise against its own classified systems
→
Mythos breaches nearly all of them within hours
→
Commerce Dept. restricts foreign access; Anthropic halts both models
Game changer
An AI that autonomously discovers vulnerabilities and chains attacks reshapes both cyber offense and defense. The NSA's own tool breaking its own systems is read as pressure on every org to harden defenses.
Skeptics say
Maybe the test environment was simply weak — or it's marketing hype, with other vendors' models yielding similar results. Hard to tell if the capability is real or the defenders fell short.
Mythos at a glance
First release
Mythos Preview · early Apr 2026
Latest version
Mythos 5 · Jun 9 (halted)
Availability
Project Glasswing partners only
Reference pricing
$10 / $50 per M tokens (in / out)
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…