OpenAI Releases GPT-5.5-Cyber to Help Defenders Find and Patch Vulnerabilities Faster

OpenAI has released the full version of GPT-5.5-Cyber, a specialized cybersecurity model built to help verified defenders find vulnerabilities, validate exploitability, generate patches, and prepare remediation evidence inside controlled workflows.

The launch expands OpenAI Daybreak, the company’s cybersecurity initiative focused on moving defenders beyond vulnerability discovery and toward faster patching. OpenAI says the model will remain limited to trusted defenders whose work requires advanced cyber capabilities and stronger oversight.

The company is positioning GPT-5.5-Cyber as a higher-capability tier for specialized authorized work. For most organizations, OpenAI Daybreak still recommends GPT-5.5 with Trusted Access for Cyber and Codex Security as the starting point for common defensive workflows.

GPT-5.5-Cyber Raises OpenAI’s Cybersecurity Benchmark Scores

According to OpenAI’s Daybreak announcement, GPT-5.5-Cyber reached 85.6% on CyberGym in single-model evaluations, compared with 81.8% for GPT-5.5. OpenAI says this was the highest single-model CyberGym score it has measured.

The model also scored 39.5% on ExploitGym, up from 25.95% for GPT-5.5. ExploitGym tests whether agents can turn known vulnerabilities into working exploits that achieve unauthorized code execution in evaluation settings.

On SEC-bench Pro, which evaluates long-horizon vulnerability discovery and proof-of-concept generation across complex software targets, GPT-5.5-Cyber reached 69.8%, compared with 63.1% for GPT-5.5.

Benchmark	GPT-5.5-Cyber	GPT-5.5	What it measures
CyberGym	85.6%	81.8%	Reproducing known vulnerabilities in software environments
ExploitGym	39.5%	25.95%	Generating working exploit behavior from known vulnerabilities in test settings
SEC-bench Pro	69.8%	63.1%	Long-horizon vulnerability discovery across complex targets

What GPT-5.5-Cyber Can Do

OpenAI says GPT-5.5-Cyber can sustain deeper analysis across large codebases. It can identify security-relevant components, trace whether vulnerable code is reachable, validate likely issues in controlled environments, develop and test patches, and prepare evidence for human review.

The key change is not just faster bug finding. OpenAI says the goal is to help defenders move through the full remediation loop, from discovery and validation to patch development and proof that a fix works.

That distinction matters because security teams often face large backlogs of vulnerabilities. A tool that produces more findings without helping teams land fixes can increase pressure on developers and maintainers.

• Map vulnerable code paths across large repositories.
• Validate whether reported issues are reachable and exploitable.
• Generate patch suggestions for human review.
• Prepare evidence that helps teams verify remediation.
• Support authorized red teaming, penetration testing, and exploit validation under stronger controls.

Codex Security Gets a Bigger Role in Patching

OpenAI is also updating the Codex Security plugin, which supports defensive code scanning and remediation workflows. The company says Codex Security has scanned more than 30 million commits across more than 30,000 codebases since its March research preview.

OpenAI says human reviewers have marked more than 70,000 findings as fixed, while more than 500,000 findings have automatically been determined to be fixed. The updated plugin can run deep scans, review recent changes, generate reports, trace attack paths, validate findings, and create codebase-specific patches for review.

The original Codex Security research preview described the tool as an application security agent that builds context about a project, creates threat models, prioritizes issues, validates findings, and proposes fixes with system-specific context.

Feature	How it helps defenders
Deep codebase scans	Reviews large repositories or selected code areas for security issues
Attack path tracing	Shows whether vulnerable code can be reached in realistic conditions
Patch generation	Creates proposed fixes that developers can review before merging
SARIF and CodeQL support	Helps teams connect findings to existing security and vulnerability management tools

Access Remains Limited to Verified Defenders

GPT-5.5-Cyber is not being released as a standard consumer or developer model. OpenAI says it is intended for verified defenders whose authorized work requires more advanced cyber capabilities and more permissive behavior, paired with verification, monitoring, scoped controls, and review.

The company’s Trusted Access for Cyber overview says GPT-5.5 with Trusted Access can support many defensive workflows, while GPT-5.5-Cyber is meant for narrower specialized work such as red teaming, exploit validation, exploit development, penetration testing, and threat hunting.

The same OpenAI help page says approval for Trusted Access does not automatically include GPT-5.5-Cyber. Access may require additional approval and controls, and OpenAI says users must stay within lawful and authorized cybersecurity work.

Patch the Planet Targets Open Source Security

OpenAI also announced Patch the Planet, a Daybreak initiative built with Trail of Bits to help open-source maintainers validate and fix security issues. The program also includes collaboration with HackerOne and Calif.

Initial participating projects include cURL, NATS Server, pyca/cryptography, Sigstore, aiohttp, Go, freenginx, Python, and python.org. These projects support widely used networking, cryptography, software supply chain, and language infrastructure.

The Patch the Planet initiative gives participating projects access to ChatGPT Pro, conditional access to Codex Security, and API credits for open-source development, maintainer automation, and release workflows.

• Trail of Bits security engineers work directly with maintainers.
• Researchers validate findings before they reach project teams.
• Maintainers remain in control of patches and disclosure processes.
• The program focuses on fixes, tests, reusable workflows, and coordinated disclosure.

Daybreak Brings Models, Partners, and Government Work Together

OpenAI says Daybreak brings together frontier cyber models, Codex Security workflows, trusted access controls, ecosystem partners, open-source maintainers, and critical infrastructure operators.

The broader Daybreak program is built around authorization, human judgment, monitoring, safeguards, and collaboration with security organizations. OpenAI also says it has established Trusted Access for Cyber partnerships with Australia, Canada, France, Germany, Japan, South Korea, and EU institutions including ENISA.

OpenAI also says it has continued collaboration with the Center for AI Standards and Innovation on pre-deployment testing for GPT-5.5 and GPT-5.5-Cyber, and work with the Office of the National Cyber Director and the Office of Science and Technology Policy on AI security implementation.

Why This Matters for Cybersecurity Teams

The launch shows how quickly AI-assisted security is moving from finding bugs to helping teams fix them. That shift could be important for enterprises, software vendors, open-source maintainers, and government agencies that already struggle to keep up with vulnerability backlogs.

At the same time, GPT-5.5-Cyber’s limited access model shows that OpenAI sees advanced cyber automation as a high-risk capability. The company is keeping the most permissive workflows behind verification, monitoring, account-level controls, and human review.

For most organizations, the practical starting point is not GPT-5.5-Cyber itself. It is Codex Security, GPT-5.5 with Trusted Access for Cyber, and better integration between vulnerability discovery, patch generation, developer review, and security validation.

FAQ