DayBreak doesn't make your systems vulnerable

Since OpenAI put Codex Security into research preview in March, the tool has scanned more than 30 million commits across over 30,000 codebases. Human reviewers marked more than 70,000 findings as fixed, and another 500,000-plus were automatically determined to be fixed. OpenAI is now folding that work, a cyber-tuned model, and two partner programs into a bundle it calls Daybreak.

The pitch rests on one claim: AI moved the hard part of security. Finding serious vulnerabilities used to take rare expertise and time inside a complex codebase. Models now read across large repositories, reason through attack paths, validate hypotheses, and surface issues fast enough that defenders drown in them. The constraint moves to patching, because a vulnerability report protects nobody on its own. The work that matters comes after: validating the issue, judging impact, building and testing a fix, coordinating disclosure, and getting the patch deployed. Daybreak is OpenAI’s attempt to automate those later steps instead of only the discovery at the front.

The accessible piece is the updated Codex Security plugin. It runs deep scans or reviews recent changes, then produces reports with severity, affected code locations, validation evidence, and remediation guidance. It traces attack paths, builds a threat model (or generates one when a team has none), checks whether vulnerable code is actually reachable, writes a codebase-specific patch, and verifies the result. It can also ingest existing findings from scanners, advisories, bug-bounty reports, or ticketing systems, triage them, and generate patches at scale to clear a backlog. Output exports to vulnerability-management systems and to SARIF files and CodeQL queries, and it runs through the Codex CLI or the Codex app. A human decides which findings to chase, which patches to apply, and what to disclose.

The model underneath ships in two tiers. GPT-5.5-Cyber, now released in full after a permissive-only preview, posts 85.6% on CyberGym against 81.8% for GPT-5.5, which OpenAI calls its highest single-model CyberGym score. It also reaches 69.8% on SEC-bench Pro, a long-horizon discovery and proof-of-concept benchmark, versus 63.1%. The figure worth sitting with is ExploitGym, which measures whether an agent can turn a known vulnerability into a working exploit that achieves unauthorized code execution: GPT-5.5-Cyber scores 39.5%, up from 25.95%. The same model that writes patches is benchmarked on writing exploits.

The same weights do defence and offence; what separates the two is who can run them. OpenAI restricts GPT-5.5-Cyber to “verified defenders” with extra verification, monitoring, scoped controls, and review, and tells most teams to stay on GPT-5.5 with Trusted Access for Cyber plus Codex Security. Early Daybreak work has turned up vulnerabilities in Firefox, V8, Safari, OpenBSD, FreeBSD, and HTTP/2 implementations.

A separate Daybreak Cyber Partner Program lets security vendors embed GPT-5.5 with Trusted Access inside their own products, so their customers get the capability without direct model access.

The open-source arm, Patch the Planet, is founded with Trail of Bits in collaboration with HackerOne and Calif. It funds security researchers, equips them with Codex Security, and sets them to work alongside maintainers. The problem is specific: faster discovery means more reports, many of them low-quality false positives, landing on teams with no spare capacity. Linux Foundation and Harvard research cited in the announcement found that 94 percent of the widely used projects it studied had fewer than ten developers responsible for over 90 percent of the code added in a year. Patch the Planet’s answer is to have researchers validate and deduplicate both the vulnerabilities and the patches before anything reaches a maintainer, with the maintainer setting priorities and disclosure process up front. Participating projects get ChatGPT Pro, conditional Codex Security access, and API credits. More than 30 projects have committed, including cURL, Go, Python, Sigstore, and pyca/cryptography. A five-day opening sprint surfaced hundreds of issues, merged dozens of patches, and left behind reusable fuzzing, variant-analysis, differential-testing, and specification-based testing workflows.

If you run systems and want to use any of this now, the Codex Security plugin on code you own is the entry point: scan it, feed it your existing finding backlog, let it draft patches, export through SARIF or CodeQL into whatever you already run, and keep a reviewer on what merges. If you maintain a widely used project, Patch the Planet is the route that hands you validated patches rather than raw report volume. GPT-5.5-Cyber itself needs verification most teams will not clear.

Two things are worth tracking. In the past month OpenAI has signed Trusted Access for Cyber arrangements with Australia, Canada, France, Germany, Japan, South Korea, EU institutions including ENISA, and the UK, alongside pre-deployment testing with CAISI and Executive Order implementation work with ONCD and OSTP. And it says it plans to fold “broader context and identifiers about the specific systems” customers operate into its safeguards. Strengthening abuse prevention that way means handing OpenAI an inventory of the infrastructure you are defending, which is its own thing to weigh.