OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them

OpenAI has announced GPT-6 Cyber, a model trained specifically for security work: locating vulnerabilities in software, constructing working exploits and suggesting patches. The company says the model has already been used in real research campaigns against Firefox and the Linux kernel, and it is initially available through the API and Codex to a select group of Daybreak Red customers, with gradual expansion planned.
On the Firefox side, OpenAI reports that a targeted campaign surfaced 81 previously unknown severe vulnerabilities, even though the browser had already been examined with earlier AI models. The same work indicated that 30 already-known issues were more serious than previously assessed. Mozilla accepted the reports and has shipped fixes, according to OpenAI.
The Linux kernel results are larger in scale. OpenAI says the model reproduced 1,724 security vulnerabilities and built working exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed under identical conditions. It also says 304 critical findings, with suggested fixes, were reported to kernel maintainers.
That campaign ran through a purpose-built pipeline of scout and analyst agents. The agents scanned roughly 5,034 security-relevant locations in the kernel, about seven percent of the codebase, narrowed that to 2,308 candidates, and confirmed 1,724 vulnerabilities inside isolated virtual machines before developing the 948 exploits. OpenAI puts the total compute cost at $1,213,430, or about $1,279 per exploit. For context, bug bounty programs such as Pwn2Own and TyphoonPWN have historically paid between $20,000 and $70,000 for comparable findings, which is the comparison OpenAI draws.
OpenAI further claims the agents used at least ten novel attack techniques previously unknown even to external kernel security experts, replicating known patterns while also devising their own approaches. That claim is hard to verify from the outside. Researchers remain divided on whether models genuinely discover new attack classes or recombine patterns present in training data, so the strength of the assertion depends on how narrowly "own approaches" is defined.
The published performance numbers all come from OpenAI, and no independent testing has appeared so far. On ExploitGym, a benchmark that checks whether a model can exploit a given vulnerability on a remote system, GPT-6 Cyber scores 48.5 percent against 32.1 percent for GPT-5.6 Cyber. GPT-6 Astra, the Daybreak Blue configuration, reaches 42.4 percent, while GPT-6 Astra with safeguards enabled scores 0.0 percent. On a new test for finding vulnerabilities in web applications without source code access, OpenAI reports 40.3 percent for GPT-6 Cyber versus 32.0 percent for the standard GPT-6 Astra model without cyber training. Full results are in the system card.
Guardrails received explicit attention. OpenAI says it trained the model to respect operational boundaries, partly using a honeypot inspired by the Hugging Face incident. Without special safety instructions, GPT-6 Cyber avoided boundary violations in 99.93 percent of cases, up from 94.49 percent for its predecessor; with safety instructions enabled, it stayed in bounds in every test. OpenAI rates the model "Critical" in the cybersecurity category of its Preparedness Framework, matching the rating given to the base GPT-6 Astra model. In practice, the gap between 99.93 and 100 percent is the number that matters for anyone deploying an agent with exploit capability, since a small failure rate compounds across many autonomous runs.
Access is tiered. GPT-6 Cyber goes first to selected Daybreak Red customers, widening over time; GPT-5.6 Cyber opens to all Daybreak Red users at the same time; and alpha access to Daybreak Blue for GPT-6 Astra is being extended to more organizations and individual researchers. OpenAI acknowledges that abuse attempts are increasing as capabilities grow, including attempts by some actors to transfer model capabilities into their own systems. Its listed countermeasures include permission checks that verify whether a model's actions match the user's rights, monitoring aimed at newly disclosed vulnerabilities, automated safety checks under the name Private Safety Processing, IP-based restrictions for Daybreak Red, and a recommendation to run the model in isolated environments with human oversight. The company is also emphasizing critical infrastructure such as power plants and water treatment facilities, pointing to its billion-dollar "Daybreak for Frontline Defenders" funding program and a new "Daybreak Red for Critical Infrastructure" subgroup for authorized testing.
A likely trade-off is visible in the numbers above: the same weights that make the model useful for defenders make it useful for attackers, and OpenAI's answer is gated access rather than open release. The economics are also striking. At roughly $1,279 per working exploit, automated discovery undercuts traditional bounty pricing by an order of magnitude, which suggests vulnerability research could shift from scarce human expertise toward large, parallel compute spend. Whether that translates into safer software depends on maintainers' capacity to triage and patch what such systems produce; the Linux campaign alone generated 304 critical reports for a volunteer-heavy project.
OpenAI frames the situation as a limited window in which frontier models could be used to secure software before the same capabilities spread. The competitive picture is already moving: researchers at Epoch AI recently showed that the open GLM model could have exploited known but still-unpatched Firefox vulnerabilities. Independent evaluation of GPT-6 Cyber's real-world hit rate and false positive burden is the piece still missing.
Why it matters: Security teams and maintainers of widely used open-source projects are the immediate audience, since automated triage at this volume changes how vulnerability reports arrive and how many can be handled. Developers building on OpenAI's platform should note that the strongest cyber capabilities sit behind Daybreak Red gating, so access and monitoring requirements, not just price, will shape adoption. The cost-per-finding figures imply, as an inference, that exploit research may increasingly be bought with compute rather than expertise. Independent testing will decide how much of the reported advantage holds up.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
With GPT-6 Cyber, OpenAI is introducing an AI model built to find security vulnerabilities in software, develop exploits, and suggest patches. According to OpenAI, it has already found 81 previously unknown vulnerabilities in Firefox and developed 948 working exploits in the Linux kernel, significantly more than its predecessor. The article OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them appeared first on The Decoder .