Breaking

OpenAI Unveils Updated GPT-5.5-Cyber, Hitting 85.6% on CyberGym

June 22, 2026 at 17:02 EDT

Foundation Models
Security
Software Dev & Coding

OpenAI announced an updated version of its cyber-defense-focused model "GPT-5.5-Cyber" around June 22, 2026. It scored 85.6% on the CyberGym benchmark, which measures vulnerability-reproduction ability, up from 81.9% in its early version, and also released a new Security plugin for its Codex development tool.

June 22 · OpenAI Daybreak

From finding flaws to fixing them at machine speed

OpenAI releases the full GPT-5.5-Cyber model and a Codex Security plugin — an agentic workflow that scans, verifies, patches and prepares evidence, aiming to put vulnerability fixes in defenders' hands.

CyberGym benchmark — vulnerability reproduction

85.6%

GPT-5.5
Cyber

81.8%

GPT-5.5
base

1,507

vulnerabilities in the CyberGym evaluation

188

projects tested for PoC generation

The agentic AppSec loop

Scan

→

Verify

→

Generate patch

→

Evidence for human review

GPT-5.5-Cyber (full)

Limited release for vetted defenders of critical infrastructure — top tier of Trusted Access for Cyber, with added verification, scoping, logging and controls.

Codex Security plugin

Available to Codex users — traces, verifies and fixes vulnerabilities, plus red teaming, pen testing and exploit validation within authorized scope.

The promise

Solved multi-step attack simulations end-to-end at a level comparable to Claude Mythos; partners report faster case resolution. Framed as democratizing fixes in defenders' hands.

The caution

Full exploit chains against the hardest targets remain unstable, real-world evaluation of the new features is limited, and the gated approach leaves lingering misuse concerns.

Continue reading

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