Breaking

Anthropic Withholds Frontier Model Claude Mythos as Cyber Skills Trigger Export Controls

June 17, 2026 at 04:46 EDT

Foundation Models
Security
Policy & Regulation

Claude Mythos, the frontier model Anthropic unveiled on April 7, 2026, displayed outsized cybersecurity capabilities as a byproduct of stronger software-engineering and reasoning skills, and the company declined to release it publicly over misuse risks. Around June 12 it was temporarily suspended under a U.S. export-control directive.

April–June 2026 · Anthropic

A Frontier Model Too Dangerous to Release Freely

Claude Mythos can autonomously discover zero-day vulnerabilities and build working exploits — a capability flagged as a national security risk. Instead of a public launch, Anthropic restricts it to defensive use through "Project Glasswing."

~2×

ExploitBench (cyber) capability vs Claude Opus 4.8

Token context window at the Preview stage

1,000s

High- & critical-severity flaws found by partners

The Watershed: Exploit-Writing Capability

ExploitBench (cyber) — Mythos roughly doubles the previous Opus class.

~39%

Claude Opus 4.8

78.0%

Claude Mythos

What's new isn't finding flaws — it's autonomously building the exploit.

Benchmark Scores

SWE-Bench Pro · agentic coding80.3%

ExploitBench · cyber78.0%

BioMysteryBench · biology46.1%

Project Glasswing — The Defensive Flow

Restricted to defenders, ahead of attackers

Partner access only
AWS · Google · Microsoft · CrowdStrike

→

Autonomous discovery
Zero-day flaws in critical software

→

Patch before exploit
Thousands of severe flaws fixed

The Case For

Defenders gain a head start, uncovering high- and critical-severity vulnerabilities in critical software before attackers can weaponize them. Mythos 5 ships alongside "Fable 5," a hardened safety variant.

The Concern

Guardrail flaws and jailbreaks let even non-experts generate sophisticated attack code. U.S. export controls forced temporary suspension of Fable 5 and Mythos 5 — broad public access is set to stay highly restricted.

Continue reading

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