Breaking

OpenAI's GPT-5.5-Cyber Hits 85.6% on CyberGym, Tops Mythos

June 22, 2026 at 14:05 EDT

Security
Foundation Models
AI Agents

Around June 22, 2026, OpenAI announced an expansion of its "Daybreak" effort, including a full version of its defense-focused "GPT-5.5-Cyber" model, and said the model surpassed Anthropic's "Claude Mythos" on CyberGym, a benchmark built around real-world software vulnerabilities.

Cybersecurity · OpenAI

A Sharper Bug-Hunting AI — Behind a Locked Door

OpenAI's upgraded, exploit-capable cyber model now matches or beats the "Mythos" model at finding real software vulnerabilities — and stays restricted to a vetted set of "verified defenders."

CyberGym

The benchmark in play — measures finding real bugs in real software

≥ Mythos

New model is positioned to approach or surpass the comparison model

Verified

Access limited to designated "verified defenders" only

CyberGym — Real-Bug Discovery

Mythos

prior best

OpenAI (new)

matches → surpasses

The upgraded model reaches at least Mythos-level performance — and edges past it — on real-world vulnerability discovery.

The Dual-Use Problem

The same skill that hardens software can be turned to attack it.

Find & fix bugs (defense)

← same capability →

Build exploits (offense)

Why it matters

Stronger results on a real-bug benchmark signal that AI-assisted vulnerability research is becoming genuinely practical for defenders.

Open questions

How far should access to an exploit-capable model extend? Specs and any wider availability remain undisclosed.

Continue reading

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