← Back to feed

OpenAI has expanded Daybreak, its AI-assisted security programme, with three simultaneous moves: a broader release of GPT-5.5-Cyber, an updated Codex Security plugin, and a new Patch the Planet initiative targeting critical open-source infrastructure.

GPT-5.5-Cyber: 85.6% CyberGym

GPT-5.5-Cyber moves out of its initial permissive-only preview into a full release through OpenAI’s limited, trusted-defender access track. On CyberGym, the dedicated cyber-evaluation benchmark, the model scores 85.6% against 81.8% for the standard GPT-5.5 — a 3.8-point gap that widens further on multi-step exploitation and remediation chains.

The model is not public API; distribution is through Daybreak partners and the Cyber Partner Program. OpenAI frames this as deliberate: “Frontier defensive capabilities should not be concentrated in the hands of a few” — but the converse of that, giving offensive capability to the wrong parties, is the constraint on access.

Codex Security

The Codex Security plugin receives an update incorporating learnings from internal usage and customer deployments. The revised version targets two distinct problems: accelerating discovery and patching of vulnerabilities in existing production systems, and preventing new vulnerabilities from reaching production in the first place. OpenAI describes the second function as the higher-value one now that AI has made the initial discovery step cheap.

Patch the Planet

The headline initiative is Patch the Planet, built with Trail of Bits. The programme pairs AI-assisted security research using frontier models with expert human security engineers, who review findings before they reach maintainers, develop patches and tests alongside project teams, and build reusable security workflows that persist after the initial engagement closes.

Additional coordination partners: HackerOne for vulnerability triage and coordinated disclosure; Calif for focused vulnerability discovery.

Initial participating projects: cURL, NATS Server, pyca/cryptography, Sigstore, aiohttp, the Go project, freenginx, Python, and python.org. The selection concentrates on widely used networking, cryptography, software supply chain, and language runtime infrastructure — components with broad downstream blast radius where a single patched flaw protects many downstream users simultaneously.

The process is opt-in and maintainer-led. Each engagement begins with a scoping consultation to understand the project’s needs and where security effort is most useful — whether vulnerability validation, patch development, CI/CD hardening, or longer-term security engineering.

The Bottleneck Has Shifted

OpenAI’s framing for the whole expansion carries a specific claim about the current state of security: “AI has changed the physics of cybersecurity.” Finding serious vulnerabilities required rare expertise and time; models can now navigate large codebases, reason through attack paths, and surface issues at scale.

The implication is that the bottleneck has moved. Historically it was finding vulnerabilities; now defenders are processing more reports with the same resources. Patch the Planet is explicitly designed for the post-discovery problem — converting findings into shipped fixes, not just generating more vulnerability reports.

It is worth noting that “Glasswing Month One” earlier this year found similar dynamics on the offensive side: AI-automated discovery produces far more candidates than human teams can triage and patch. OpenAI and Anthropic are now both running programmes that try to close that gap on the defensive side.

Key Numbers

Metric	GPT-5.5	GPT-5.5-Cyber
CyberGym	81.8%	85.6%
Access	Public API	Trusted defenders only

Patch the Planet initial cohort: 9 open-source projects. Additional projects will join in future rounds.