Breaking

OpenAI Opens Upgraded GPT-5.5-Cyber Model to Vetted Defenders

June 22, 2026 at 13:13 EDT

Security
Foundation Models
AI Agents

On June 22, OpenAI began a limited rollout of a more capable, upgraded version of GPT-5.5-Cyber, a model fine-tuned for cybersecurity, to vetted security firms and researchers. It also announced an expansion of its defensive AI program, OpenAI Daybreak.

June 22 · OpenAI

A Sharper Cyber Model — for Defenders Only

OpenAI is shipping an updated, less-restrictive GPT-5.5-Cyber for vetted security teams — deeper codebase analysis, vulnerability validation, and patch testing, delivered through its trust-based Trusted Access for Cyber program rather than to the public.

85.6%

CyberGym score, updated Cyber model

3,000+

Critical/high vulnerabilities auto-remediated via Codex Security

National & EU partners with vetted access

CyberGym Benchmark: Updated vs. Base

Can the AI agent reproduce known software vulnerabilities? Higher is better.

81.8%

GPT-5.5 (base)

85.6%

GPT-5.5-Cyber (updated)

The Road to Trusted Access

2025–26
GPT-5.2 cyber-safety training begins

→

Feb 2026
TAC program launched

→

Apr–May
GPT-5.4-Cyber ships; 5.5 preview

→

Jun 22
Updated GPT-5.5-Cyber rolls out

The case for broad defender access

Agentic coding speeds vulnerability discovery and remediation. Getting capabilities to defenders early — secure code review, triage, malware and binary analysis, patch validation — is seen as key before attackers exploit the same power.

The open questions

It remains unverified whether it clearly beats rival Claude Mythos. With access limited to vetted participants, real-world experience beyond the benchmark is still scarce — and the dual-use risk must be managed as capability grows.

Continue reading

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