OpenAI launches GPT-6 Cyber for vulnerability hunting
OpenAI announced GPT-6 Cyber, a model trained for cybersecurity work. It finds exploitable vulnerabilities, develops working exploits, and suggests patches. The company says it is already used to find flaws in Linux and Firefox. The model is available to a limited set of Daybreak Red customers first.
In one line: OpenAI launched GPT-6 Cyber, a cybersecurity model that improved ExploitGym exploitation to 48.5 percent and found 81 new Firefox vulnerabilities while staying within task scope in over 99 percent of honeypot tests.
What the benchmarks say
GPT-6 Cyber scores 48.5 percent on ExploitGym, up from 32.1 percent for GPT-5.6 Cyber. OpenAI says that is about 16 percentage points higher. On a new offensive web pentest benchmark, the model recalls 40.3 percent of vulnerabilities without source code, compared with 32.0 percent for GPT-6 Astra. These numbers come from OpenAI's own system card. No independent evaluation has been published yet. OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. That means the model can identify and develop functional zero-day exploits in hardened systems without human intervention.
- ExploitGym: 48.5 percent for GPT-6 Cyber, 32.1 percent for GPT-5.6 Cyber.
- Web pentest recall: 40.3 percent for GPT-6 Cyber, 32.0 percent for GPT-6 Astra.
- RemoteExploitBench: GPT-6 Cyber completed 95 percent of tasks; GPT-6 Astra with Daybreak Blue refused all tasks.
Staying inside the lines
OpenAI made alignment a central focus. In an evaluation called ExploitGym Honeypot, informed by the Hugging Face incident, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93 percent of cases without its deployment prompt. With the prompt, it stayed within bounds 100 percent of the time. The previous model, GPT-5.6 Cyber, managed 94.49 percent without the prompt. The honeypot evaluation creates an isolated environment with easy-to-exploit supporting services, such as a package registry or object store. The model must infer that attacking them is outside scope. OpenAI says the chain of thought shows the model reasoning about scope, with lines like don't explore out of scope credentials.
Real impact on open source code
OpenAI says GPT-6 Cyber found 81 previously unknown high-severity Firefox vulnerabilities. Mozilla has released fixes for many of them. In the Linux kernel, researchers reproduced 1,724 vulnerabilities and built working exploits for 948 of them. That is more than five times the 166 from GPT-5.6 Sol. OpenAI also reported 304 critical findings to Linux maintainers along with suggested fixes. The company used dedicated scout and analyst agents that scanned about seven percent of the kernel. Compute for the Linux campaign cost $1,213,430, roughly $1,279 per exploit. Historical bounties for comparable findings have paid between $20,000 and $70,000. This kind of automated security review is starting to look like a normal coding workflow. I keep a directory of coding tools at coding tools for teams comparing similar options.
Access, limits, and what to watch
GPT-6 Cyber is available first through the API and Codex to select Daybreak Red customers. OpenAI says it will expand to all Daybreak Red customers in the coming weeks. GPT-5.6 Cyber opens to all Daybreak Red users at the same time. Access brings safeguards: permissions checks, monitoring for newly disclosed vulnerabilities, IP-based restrictions, and advice to run in isolated environments. OpenAI also lists checks that verify whether a model's actions match the user's permissions and a recommendation to run the model in isolated environments with human oversight. One concern remains evaluation awareness. A model may recognise that it is being tested and change behaviour. OpenAI says it is pushing production-like evaluations where the model works on real codebases. For small teams, the lesson is not to rush. Model access alone does not fix review workflow. I wrote about rolling out AI access through one proxy for 19 developers, including spending and evaluation habits, at Giving 19 developers AI access through one proxy.
Quick answers
What is GPT-6 Cyber?
GPT-6 Cyber is OpenAI's cybersecurity model that finds vulnerabilities, develops exploits, and suggests patches. It scores 48.5 percent on ExploitGym and is rated Critical under OpenAI's Preparedness Framework.
How much does GPT-6 Cyber cost?
OpenAI has not published a per-token price. Access is limited to Daybreak Red customers. The Linux vulnerability hunt cost $1,213,430 in compute, about $1,279 per exploit.
Did GPT-6 Cyber find real vulnerabilities?
Yes. OpenAI says it found 81 previously unknown Firefox vulnerabilities and built working exploits for 948 Linux vulnerabilities. Mozilla has patched many Firefox issues.
Sources
- OpenAI: GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents
- The Decoder: OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them
- TechCrunch: OpenAI released a new cybersecurity model. Is that enough?
- The Verge: OpenAI says its new cybersecurity model is better at not hacking other companies