Breaking

Anthropic's Mythos Flagged Flaws in Classified US Systems in Hours

June 23, 2026 at 23:02 EDT

Foundation Models
Security
Policy & Regulation

Anthropic's AI model "Mythos" identified vulnerabilities in highly sensitive US government computer systems within hours during a joint testing exercise with US intelligence agencies, a US official told the AP. The model found the vulnerabilities but did not actually exploit them.

June 23, 2026 · Anthropic × US Intelligence

AI model "Mythos" cracked classified US systems in hours, not weeks

During a joint test under "Project Glasswing," Anthropic's frontier model identified multiple vulnerabilities in high-security government systems within hours — though it did not actually exploit them.

Hours

Time to penetrate nearly all classified systems tested

244+

Pages in the System Card documenting its capabilities

100+

Cyber experts (plus Adobe & Nvidia) opposing the access curbs

Speed of vulnerability discovery

"Hours, not weeks"

Weeks

Traditional approach

Hours

Mythos in the joint test

A task that once took weeks of expert effort collapsed into a few hours.

Claude Mythos at a glance

Announced

April 7, 2026 (Preview)

Availability

No public release — Glasswing & select customers only

Capability

Autonomous zero-day discovery & exploit PoCs

Verdict

Judged "too dangerous to release publicly"

The case for it

Unmatched scale and speed of vulnerability discovery — a powerful defensive tool to protect critical infrastructure before attackers strike.

The case against the curbs

100+ experts, Adobe & Nvidia call to withdraw export controls: Mythos is "excellent but not unique," and restricting the best defense lets adversaries catch up.

The open question: how to balance dual-use risk — accelerating defense while arming potential attackers — against national-security access restrictions.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.