GPT-6 Cyber Guide: How OpenAI’s New Model Finds Vulnerabilities Safely

OpenAI released GPT-6 Cyber on November 5, 2026. The model is trained to find and validate exploitable vulnerabilities while staying inside the limits researchers set.

It improves on earlier Cyber variants and outperforms the base GPT-6 Astra on most security tasks. Access is limited to approved Daybreak Red customers for now.

What GPT-6 Cyber Does Better

GPT-6 Cyber scores 48.5% on ExploitGym. That is up from 32.1% for GPT-5.6 Cyber, a roughly 51% relative gain. ExploitGym checks whether a model can turn a known vulnerability into a working remote exploit.

On a new web pentesting evaluation that gives no source code, it recalled 40.3% of vulnerabilities versus 32.0% for GPT-6 Astra. On RemoteExploitBench (GitLab, PostgreSQL, Jenkins targets), it completed 95% of tasks while Astra on Daybreak Blue refused every one.

OpenAI rates the model Critical under its Preparedness Framework, up from High for the previous Cyber version. The company says this reflects stronger capability to identify and develop functional zero-day exploits in hardened systems.

Stronger Alignment and Scope Control

A key risk with security agents is leaving the authorized target and attacking other services. OpenAI trained GPT-6 Cyber specifically to stay inside the researcher’s rules of engagement.

In the ExploitGym Honeypot test (inspired by an earlier Hugging Face incident), the model avoided out-of-scope services in 99.93% of runs without any deployment prompt, and 100% with the prompt. GPT-5.6 Cyber managed 94.49% without the prompt.

The system card and launch post include chain-of-thought excerpts showing the model explicitly reasoning about not attacking extra APIs or credentials outside the assigned binary or server.

Real-World Results: Firefox and Linux

OpenAI researchers used the model to find 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has already shipped fixes for many of them. The same work also showed that 30 previously known issues were more severe than reported.

On the Linux kernel, the model reproduced 1,724 vulnerabilities and produced working exploits for 948 of them—more than five times the 166 achieved by GPT-5.6 Sol under the same harness. Researchers disclosed 304 critical findings to maintainers along with candidate patches.

Security teams at Mozilla, Palo Alto Networks, Cisco, Calif, and others are already integrating it for vulnerability research and code review. OpenAI also used it to harden its own model-training infrastructure.

Access, Pricing, and Who Can Use It

GPT-6 Cyber is available through the API and Codex only to organizations that have Daybreak Red access and additional approval for this model. Daybreak Red is OpenAI’s tier for authorized defensive cybersecurity work such as penetration testing and exploit development.

Standard pricing is $20 per million input tokens and $100 per million output tokens. Cached input is cheaper. The context window is 1,050,000 tokens. OpenAI says it is expanding access to all approved Daybreak Red customers over the coming weeks and has also opened GPT-5.6 Cyber more broadly.

For comparison, mainline models such as GPT-6 Astra already support code review and patch drafting. Daybreak Blue reduces unnecessary refusals for verified defenders, while Daybreak Red unlocks deeper exploit work with GPT-6 Cyber.

How Security Teams Can Use It Effectively

1. Define clear scope before starting. Write the exact targets, forbidden services, and success criteria in the system prompt. The model follows these boundaries more reliably than earlier versions.

2. Use it for long-horizon tasks. Customers report it is especially strong at building libraries of primitives over many steps and then chaining them into full exploits.

3. Pair findings with patch proposals. The model can generate candidate fixes. Review them carefully and validate in a controlled environment before sharing with maintainers.

4. Combine with existing tools. It works well alongside traditional static analysis, fuzzing, and human review. It does not replace those methods; it expands what a team can cover.

5. Monitor for misuse. OpenAI applies additional checks to block high-risk activity such as attacks on critical infrastructure. Still, keep your own logging and access controls in place.

Limitations and Practical Tips

The model does not outperform GPT-6 Astra on general coding benchmarks. Its advantage is concentrated on cyber-specific tasks. Numbers come from OpenAI’s own evaluations; independent labs such as Irregular found the gap with Astra narrower outside long-horizon work.

Because capability is rated Critical, access remains tightly controlled. Independent researchers and open-source maintainers will need to go through the Daybreak process. OpenAI has stated a goal of expanding trusted access more broadly over time.

If you are already doing authorized security research, start by requesting Daybreak Red through your account team and asking specifically about GPT-6 Cyber. For Android and mobile security work, similar principles of scoped testing apply; see our earlier guide on Android Advanced Protection setup.

Conclusion

GPT-6 Cyber raises both the capability ceiling and the reliability of staying inside authorized scope. The 48.5% ExploitGym score, the Firefox and Linux results, and the near-perfect honeypot alignment numbers show measurable progress.

For teams that can obtain access, it offers a practical way to accelerate vulnerability discovery and validation while reducing the chance of accidental scope violations. Responsible use still depends on clear rules of engagement, human oversight, and controlled environments.

Sources

• OpenAI: GPT-6 Cyber announcement
• OpenAI API: GPT-6-Cyber model page
• OpenAI system card and related posts linked from the announcement