Breaking

Anthropic's frontier model Mythos reportedly breached classified systems in NSA test

June 24, 2026 at 07:14 EDT

Security
Foundation Models
Policy & Regulation

Anthropic's unreleased frontier model "Claude Mythos" reportedly discovered vulnerabilities in highly sensitive U.S. government systems during an intelligence-agency red-team test, breaching nearly all targeted systems within hours. In response, the U.S. government around June 2026 immediately suspended worldwide access to Mythos 5 and Fable 5.

June 2026 · Anthropic · Project Glasswing

Claude Mythos: A Frontier Model Built to Hunt Software Flaws

A non-public cybersecurity model unveiled in April 2026 reportedly uncovered thousands of zero-day vulnerabilities — and the contested, still-unverified claim that it breached classified U.S. systems has fueled both alarm and skepticism.

6,202+

High/critical vulnerabilities found (of 23,019 total)

1,000+

Open-source projects scanned

83%+

Working exploit on first attempt

~50

Organizations with limited access (no public release)

Bug Discovery: Partners Report 10× the Old Rate

Glasswing participants such as Cloudflare, Mozilla and Microsoft cite roughly ten times higher bug-discovery rates.

1×

Before

10×

With Mythos

Defensive wins (verified)

A bug undetected in OpenBSD for 27 years
wolfSSL certificate-forgery flaw (CVE-2026-5194)
Pentagon deploying it to patch cyber gaps
UK AISI confirmed gains in CTF and multi-step attack simulations

Caution & open questions

Classified-system intrusion claim lacks primary evidence
Lower reliability against well-hardened systems
"If threat actors use it, patching won't keep up"
Moves toward export controls and access restrictions

The dual-use problem

The same capability that finds and fixes flaws can also weaponize them — so access stays tightly restricted.

Defense-only

Glasswing protects key software

No public release

Mythos 5 updated June 9, kept private

Expansion blocked

A scale-up reportedly halted by the White House

Continue reading

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