Breaking

OpenAI Opens GPT-5.5-Cyber to Vetted Defenders Only

June 22, 2026 at 23:22 EDT

Security
Foundation Models
AI Agents

OpenAI began rolling out GPT-5.5-Cyber, a model specialized for cybersecurity, around May 7, 2026, as a limited preview restricted to vetted defender organizations. Under the Trusted Access for Cyber (TAC) program, it offers more permissive behavior than standard GPT-5.5 for authorized red teaming, penetration testing and controlled validation, while strengthening account-level controls and verification, including phishing-resistant authentication.

June 22, 2026 · OpenAI

GPT-5.5-Cyber: an AI that patches bugs, locked to verified defenders

Part of the "Daybreak" initiative, the updated model automates vulnerability work end to end — detection, patch generation and validation — but access is gated through Trusted Access for Cyber, with strict review, monitoring and phishing-resistant authentication.

85.6%

CyberGym score — a record for a single model

30+

OSS projects in "Patch the Planet" — cURL, Go, Python

50%+

of Sophos' agentic SOC work auto-resolved by GPT-5.5

Benchmark gains over the earlier GPT-5.5

Updated GPT-5.5-Cyber (orange) vs prior GPT-5.5 (gray) · column height scaled to score

85.6%

81.8%

CyberGym

39.5%

26.0%

ExploitGym

69.8%

63.1%

SEC-bench Pro

The defender workflow, automated end to end

Detect

→

Validate

→

Generate patch

→

Test & evidence

As AI gets better at finding bugs, OpenAI argues that applying patches — not discovery — is now the bottleneck.

For approved defenders

Fewer refusals and a practical, end-to-end tool. AISI found GPT-5.5 strong on cyber tasks and the second model to solve multi-step attack simulations end to end.

Open concerns

Misuse risk and unequal access. Developers also report routine devops tasks drawing more cyber-risk flags after migrating to GPT-5.5.

Continue reading

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