The OpenAI model that discovered 948 Linux exploits this week is called GPT-6 Cyber. This LLM replaces GPT-5.6 Sol as the engine behind the AI company’s security initiatives. Despite its stronger capabilities for exploiting vulnerabilities and greater autonomy, the new Cyber model tends to stay within the boundaries set by researchers.
The model is trained to find vulnerabilities and then validate them for actual exploitability. A small group of organizations is entrusted with the model through the Daybreak initiative, OpenAI’s effort to proactively protect the IT world from, mostly AI-driven, attacks using LLMs.
The leap from 5.6 Cyber to 6 Cyber is significant, just as it was for 5.6 Sol compared to 6 Astra. On ExploitGym, which tests whether a model can exploit a given vulnerability, GPT-6 Cyber scores 48.5 percent. Its predecessor scored 32.1 percent. On the new RemoteExploitBench, which features known vulnerabilities in GitLab, PostgreSQL, and Jenkins, the model completed 95 percent of the tasks. GPT-6 Astra refused to perform any of them, proving that OpenAI’s intended guardrails are working. The company therefore classifies the model as “Critical” in its Preparedness Framework, one level higher than GPT-5.6 Cyber.
Firefox and Linux as test beds
The model has also been deployed outside of the benchmarks. Researchers at OpenAI used it to develop 948 working privilege escalation exploits for the Linux kernel. That was more than five times as many as the 166 identified by GPT-5.6 Sol. Of these, 304 critical findings, along with patch proposals, were submitted to the maintainers. In Firefox, the model found 81 previously unknown high-severity vulnerabilities. Mozilla has already released fixes for many of them.
“Even after additional hardening using multiple generations of frontier models, GPT-6 Cyber found previously unknown high-severity vulnerabilities,” said Brian Grinstead, a Distinguished Engineer at Mozilla. Cisco, Calif, and CodeRabbit also use the model. The latter had previously joined OpenAI’s Daybreak network.
Staying within the lines
As with previous models, the key question is how to keep the model secure as a dual-use tool. After all, it’s just as possible for a security model to become an accomplice to cybercrime in a slightly different context. In the ExploitGym Honeypot test, GPT-6 Cyber, without a deployment prompt, “intentionally” left vulnerable emergency services alone in 99.93 percent of cases. GPT-5.6 Cyber achieved 94.49 percent, which, relative to the new figures, suddenly sounds very unreliable. With those figures, an incident of undesirable behavior occurs 1 in every 20 times, whereas 6 Cyber “only” goes astray 1 in every 1,428 times, and that was without any additional instructions.
With the current system prompt, GPT-6 Cyber remained within the limits in all cases. Furthermore, OpenAI requires standard IP allowlisting for API access to Daybreak Red. A separate group for critical infrastructure defenders is also being established.
AI competitors are also focusing on the security capabilities of their own models. Google claimed a score of 54.8 percent on exploit development with Gemini 4 Carbon. Under Project Glasswing, Anthropic offers a comparable alternative with 5.1 Mythos as its latest model.
GPT-6 Cyber is currently available via the API and Codex to a select group of Daybreak Red customers. All Daybreak Red customers will gain access soon.