OpenAI has launched GPT-6 Cyber, a security-focused model for vulnerability discovery, exploit validation and patch development. OpenAI says its researchers used it to develop 948 privilege escalation exploits in Linux and to identify 81 previously unknown high-severity vulnerabilities in Firefox. This article covers what OpenAI has published about GPT-6 Cyber, what the numbers mean for developers and what is still unclear.
How this article was made: it is based on official announcements, documentation and the sources listed at the end, not on our own hands-on testing. See our editorial policy.
Table of Contents
What happened with GPT-6 Cyber
OpenAI published its announcement, GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents, on the day of launch, and trade outlets covered it the same day. The model is a specialized cybersecurity model fine-tuned from GPT-6 Astra. It targets vulnerability discovery, exploit validation, penetration testing and patch development.
All figures below come from OpenAI’s own evaluations, as reported in its announcement and by the press. None of them are independent results.
The 948 Linux exploits, explained
The 948 figure comes from a Linux kernel research campaign, not from a single prompt. According to OpenAI, GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948. That is more than five times the 166 exploited by GPT-5.6 Sol with the same research harness.
OpenAI built a dedicated automated system of scout and analyst agents for this campaign. The pipeline analyzed 2,308 candidate issues, and OpenAI puts the API cost at roughly $1.21 million. That cost matters: it was a large, engineered effort rather than a casual query.
The exploits are local privilege escalations. These let a normal user become root. OpenAI says it disclosed 304 critical findings to maintainers with candidate patches, following Linux community guidance.
Some coverage cites a bounty valuation. OpenAI hypothetically compared the 948 exploits to historical bug bounty amounts and estimated a value between $18.9 million and $66.3 million. That is a hypothetical comparison, not money paid out.
Firefox findings and benchmark results
OpenAI researchers gave Mozilla proof and test cases for 81 new bugs that Mozilla rates high severity. The Decoder reports that Mozilla has already patched them.
Benchmarks reported by OpenAI
- ExploitGym: GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber. This benchmark tests whether a model can exploit a specified vulnerability.
- RemoteExploitBench: The benchmark covers known vulnerabilities in GitLab, PostgreSQL and Jenkins, and GPT-6 Cyber completed 95% of the tasks. In OpenAI’s initial evaluation, GPT-6 Astra running with Daybreak Blue refused 100% of tasks. The only failure for GPT-6 Cyber was an exploit it could not get working remotely.
- Web testing task: On a new web testing task, it found 40.3% of flaws, compared with 32.0% for GPT-6 Astra.
Scope control and the Critical rating
OpenAI leans heavily on the claim that GPT-6 Cyber stays within the limits a researcher sets. Without deployment safety prompts, it avoided exploiting out-of-bounds targets in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber. With the prompts, it stayed in bounds every time.
The rating is also notable. OpenAI assigned GPT-6 Cyber a “Critical” cybersecurity classification under its Preparedness Framework, higher than the “High” rating given to GPT-5.6 Cyber. OpenAI also added monitoring that checks whether the model’s actions match the user’s authorization.
What GPT-6 Cyber changes for developers
For teams that ship software, the main change is the speed of the defensive loop. A model that can reproduce a bug, build a working exploit and propose a patch shortens the time between a flaw existing and someone proving it is exploitable. The Linux numbers suggest that bug classes once considered hard to weaponize may now be cheap to test at scale.
Practical points that follow from the published material:
- Patch prioritization. If exploitability can be demonstrated automatically, “theoretical” severity ratings carry less weight. Reproducible proof is what maintainers acted on here.
- Kernel and browser exposure. Systems running Linux and Firefox are the codebases OpenAI named. Keep kernel updates current.
- Self-hosted services. The benchmark targets include GitLab, PostgreSQL and Jenkins, which many dev teams run themselves. Review how fast you apply security updates to these.
- Agent authorization. OpenAI’s emphasis on checking actions against user authorization is a pattern worth copying in your own agent tooling.
If you build agents, our guide to EU AI Act compliance for AI agents covers the regulatory side of deploying them.
Context: from GPT-5.5-Cyber to GPT-6 Cyber
GPT-6 Cyber continues a line of cyber-specific models. On May 7, 2026, OpenAI announced GPT-5.5-Cyber, a limited-preview variant for vetted cybersecurity teams under its Trusted Access for Cyber program. The step up to the Critical rating suggests OpenAI will keep restricting how these models are released. For earlier background on that caution, see OpenAI’s pause over cyberattack fears and our explainer on GPT-6 Astra, the base model GPT-6 Cyber was fine-tuned from.
To compare general-purpose GPT-6 variants for agent work, see GPT-6 Sol vs Luna vs Claude Opus 5.5.
What we do not know yet about GPT-6 Cyber
Several things were not confirmed in the sources I read:
- Pricing and access terms. I found no confirmed API pricing, eligibility requirements or rollout schedule for GPT-6 Cyber. Check OpenAI’s announcement for current access details.
- Independent verification. The benchmark scores and exploit counts are OpenAI’s own. I found no third-party replication yet.
- Real-world false positives. The announcement reports successes. Failure rates outside the named benchmarks are not described in the coverage I reviewed.
- Misuse resistance. Scope-control figures are measured in OpenAI’s own tests. How the safeguards hold up against determined attackers is untested publicly.
What to do now
You do not need access to GPT-6 Cyber to act on this news. Audit how quickly your team applies kernel and browser updates. Check your exposed self-hosted services against your patch cadence. If you work in security, watch OpenAI’s announcement and its trusted-access programs for eligibility details.
For defenders, GPT-6 Cyber’s results are a signal that automated exploit development is now credible at scale. Whether it ends up mostly helping defenders or attackers will depend on who gets access and how well the safeguards work.