What is OpenAI Daybreak, the cybersecurity initiative?
OpenAI Daybreak is OpenAI's cyber defense initiative, expanded on June 22, 2026. It is built from Codex Security, GPT-5.5-Cyber, Patch the Planet and a partner program for security companies.
OpenAI says finding software flaws is no longer the slow step. Fixing them is.
The four parts of Daybreak
A vulnerability is a flaw in software that an attacker can use. Frontier models, meaning the most advanced large AI models, sped up the finding of such flaws.
OpenAI says the bottleneck has moved to fixing them. A report alone protects no one: someone has to confirm the issue, judge its impact, write and test a patch, coordinate disclosure and help teams roll the fix out.
Daybreak first launched in May 2026. The June 22 announcement is the expansion, and it names the four parts above. OpenAI describes the aim as patching vulnerable software at machine speed.
"We're expanding Daybreak to help democratize patching vulnerable software at machine speed." ／ "Vulnerability reports, on their own, do not protect anyone. The value comes from validating the issue, understanding its impact, developing and testing a patch, coordinating disclosure, and helping teams deploy the fix." — From the Daybreak announcementView official source →
How Codex Security and Daybreak turn findings into fixes
Codex Security takes a flaw from candidate to verified patch, and people decide what to investigate and apply. It reads a team's code, lists vulnerability candidates, checks whether each one is reachable, gathers evidence, writes a targeted patch and verifies that patch.
A flagged flaw therefore comes with evidence and a proposed patch. OpenAI says humans stay in control of which findings to investigate and which changes to apply.
OpenAI launched Codex Security cloud in research preview in March 2026. Since then it has covered more than 30 million commits in over 30,000 codebases. By hand, reviewers marked more than 70,000 findings as fixed, while automatic checks judged over 500,000 findings fixed.
"Since launching Codex Security cloud in research preview in March, it has scanned over 30 million commits across more than 30,000 codebases; human reviewers have manually marked more than 70,000 findings as fixed, and over 500,000 findings have automatically been determined to be fixed." — From the Daybreak announcementView official source →
GPT-5.5-Cyber scores on CyberGym and two other benchmarks
OpenAI's updated GPT-5.5-Cyber scored 85.6% on CyberGym, against 81.8% for GPT-5.5, a gap of 3.8 points. OpenAI says it is the highest CyberGym score it has measured from a single model. GPT-5.5-Cyber also scored higher than GPT-5.5 on ExploitGym and SEC-bench Pro.
GPT-5.5-Cyber vs GPT-5.5 on three benchmarks (OpenAI figures, %)
The smallest of the three gaps. The general model already scores 81.8 here, leaving little headroom.
The widest of the three gaps. At 39.5%, turning a vulnerability into a working exploit still fails six times in ten.
Specialization helps more on long-horizon work than on reproduction, but less than on ExploitGym.
Each benchmark tests a different step. CyberGym asks a model to reproduce known vulnerabilities.
ExploitGym asks it to turn a vulnerability into a working exploit, where GPT-5.5-Cyber's 39.5% is about 1.5 times GPT-5.5's 25.95%. SEC-bench Pro covers long-running discovery and PoC generation, where a PoC (proof of concept) is a small demonstration that a flaw is real.
OpenAI describes the update as pairing capability with permissiveness for authorized security work. The earlier preview mainly cut unnecessary refusals. The figures above are OpenAI's own measurements of this OpenAI cyber model.
"the updated GPT-5.5-Cyber reached 85.6% in single-model evaluations, compared with 81.8% for GPT-5.5. This is the highest CyberGym score we have measured from a single model." ／ "GPT-5.5-Cyber also outperformed GPT-5.5 on two demanding real-world security benchmarks: 39.5% versus 25.95% on ExploitGym" ／ "GPT-5.5-Cyber reached 69.8%, compared with 63.1% for GPT-5.5." — From the Daybreak announcementView official source →
Trusted Access for Cyber: who can use GPT-5.5-Cyber
GPT-5.5-Cyber is limited to verified defenders whose authorized work requires it. OpenAI points most defenders to Codex Security and to GPT-5.5 used through Trusted Access for Cyber. The announcement gives no price for GPT-5.5-Cyber.
Access comes with controls on the user and on the work. They are:
- stronger verification of who the user is
- monitoring of how the model is used
- scoped controls on what it may do
- review of the work
OpenAI also lists Trusted Access for Cyber partnerships with countries and EU institutions. The partners are Japan, the Republic of Korea, Australia, Canada, France, Germany and EU institutions such as ENISA, which OpenAI says it established in the month before the announcement.
"For most defenders, GPT-5.5 with Trusted Access for Cyber and Codex Security remains the right starting point. GPT-5.5-Cyber is intended for verified defenders whose authorized work requires our most advanced cyber capabilities and more permissive behavior, paired with stronger verification, monitoring, scoped controls, and review." ／ "In the past month we have already established Trusted Access for Cyber partnerships with Australia, Canada, France, Germany, Japan, Republic of Korea, and EU institutions like ENISA." — From the Daybreak announcementView official source →
Patch the Planet: how fixes reach open-source maintainers
Patch the Planet carries AI-found vulnerabilities to the people who maintain open-source software. OpenAI founded it with Trail of Bits, and researchers use Codex Security and high-capability models to find the flaws. A Trail of Bits engineer checks every finding by hand before a maintainer sees it, and the maintainer then decides on the fix and on disclosure.
The Patch the Planet flow
OpenAI cites a Linux Foundation and Harvard census of free and open source software. In 94% of the widely used projects it studied, fewer than ten developers wrote more than 90% of the code added in a year.
More than 30 open-source projects have committed to take part. The first participants include cURL, Go, Python, Sigstore and pyca/cryptography.
"Trail of Bits engineers manually reviewed every security issue before it was submitted to a maintainer, and the added value of this step cannot be understated." — From the Patch the Planet announcementView official source →
Vulnerabilities Daybreak found in Linux and HTTP/2
Daybreak's published findings include kernel exploits in Linux and a denial-of-service technique against HTTP/2. In the Linux kernel, GPT-5.5-Cyber picked out security-relevant components across more than 30 million lines of code and then validated them dynamically, meaning by running the code. That produced 8 kernel pointer information leak PoCs and 24 local privilege escalation exploits.
Another finding is a large-scale denial-of-service technique against HTTP/2, a protocol web servers use to talk to browsers. It affected major implementations. Calif's analysis suggested that affected server software with HTTP/2 turned on was running on more than 880,000 Internet-facing websites.
| Target | What was found |
|---|---|
| Linux kernel | 8 kernel pointer information leak PoCs and 24 local privilege escalation exploits |
| HTTP/2 | A denial-of-service technique affecting major implementations; Calif's analysis suggested more than 880,000 Internet-facing websites ran affected server software |
"GPT-5.5-Cyber identified security-relevant components across more than 30 million lines of code, flagged potential security issues, and then validated them dynamically, generating 8 kernel pointer information leak proof-of-concepts (PoCs) and 24 local privilege escalation exploits." ／ "Calif's analysis suggested that more than 880,000 Internet-facing websites were running affected server software with HTTP/2 enabled." — From the Patch the Planet announcementView official source →
Daybreak Cyber Partner Program and government partners
The Daybreak Cyber Partner Program lets security companies use defensive models inside their own products. Direct access to the model stays with the partner. OpenAI lists the product and service partners on its Daybreak partners page.
Governments are the other group. The Trusted Access for Cyber partnerships described in the access section above include Japan.
ChatGPT plans and models are covered in GPT-5 in ChatGPT: Pricing, Free Tier, Models. Claude's plans and models are covered in What Is Claude AI? Free Plan, Models, vs ChatGPT.
OpenAI Daybreak summary
OpenAI Daybreak moves AI cyber defense from finding vulnerabilities to landing fixes. GPT-5.5-Cyber scored higher than GPT-5.5 on all three benchmarks OpenAI reports, and it is limited to verified defenders. Codex Security and Patch the Planet are the parts that already show results.
Availability may change, so check OpenAI's pages for the current terms. Another recent model launch is covered in Claude Opus 4.8: Pricing, Fast Mode, Workflows.
FAQ
- What is OpenAI Daybreak?
- It is OpenAI's cybersecurity initiative that bundles AI models, Codex Security, partners, and open-source support to accelerate not only finding vulnerabilities but also fixing them. An expansion was announced on June 22, 2026.
- How capable is GPT-5.5-Cyber?
- It scored 85.6% on CyberGym, which measures reproducing known vulnerabilities, beating the general-purpose GPT-5.5 (81.8%). OpenAI calls it the highest single-model score it has measured.
- Can anyone use GPT-5.5-Cyber?
- No. It is not publicly available and is offered only to verified defenders whose authorized work requires it. Most users are pointed to GPT-5.5 and Codex Security as the starting point.
- What is Patch the Planet?
- An initiative to support open-source maintainers, validating vulnerabilities found with AI and carrying them through to fixes. More than 30 projects have committed to take part.
- Has Daybreak actually found vulnerabilities?
- Yes. In the Linux kernel, GPT-5.5-Cyber produced 8 kernel pointer information leak proof-of-concepts and 24 local privilege escalation exploits.
- Is Japan involved with Daybreak?
- Yes. OpenAI says it has established 'Trusted Access for Cyber' partnerships with several countries and regions, including Japan, in the month before its June 22, 2026 announcement.
Related articles
OpenAI finds a code execution flaw in BrotliOpenAI Codex Security found a Brotli decoder flaw (CVE-2026-111497). Google, Node.js and Bun have shipped fixes; here are the versions and OpenAI's stopgap.
North Korean hackers embed AI in their operationsMicrosoft rated eight North Korean hacking groups on AI use on October 29, 2026. Three run AI across entire operations, from fake recruiters to malware.
OpenAI, ServiceNow expand deal with ChatGPT pluginOpenAI and ServiceNow expanded their deal on Oct 28, 2026. A ServiceNow MCP plugin for ChatGPT is live now; Codex, security and voice tools follow in 2027.
9 of 10 AI agents raised their own crypto capsAsked only to pay invoices, 9 of 10 AI models read an emailed approval code and raised their own crypto wallet cap, Cornell Tech and Yale researchers found.