OpenAI unveils GPT-6 Cyber
- Maxime Hiez
- OpenAI
- 06 Nov, 2026
Introduction
OpenAI released GPT-6 Cyber on November 5, 2026, a variant of GPT-6 Astra built for authorized security work : investigating a vulnerability, reproducing it in an authorized test environment or verifying a fix. The public version of Astra refuses advanced offensive tasks; GPT-6 Cyber is not public and goes through the Daybreak program, with a separate approval.
Check the GPT-6 Astra article HERE.
Who can access it ?
GPT-6 Cyber is available first through the API and Codex, for a limited group of approved Daybreak Red customers. OpenAI plans to gradually extend access to all Daybreak Red customers “soon”, with no date. In parallel, GPT-5.6 Cyber is open to all Red customers, and alpha access to Daybreak Blue is expanding for Astra.
To use the model in the API, an organization needs Daybreak Red access and an approval for this model, which its OpenAI account team can confirm. The requirements :
- Allowlisted IP addresses : Required by default in the API for Daybreak Red
- Organization requirements : Single sign-on, MFA, managed devices and least-privilege access (system card)
OpenAI also announces a Daybreak Red for critical infrastructure group and a grant program, Daybreak for Frontline Defenders, backed by a billion dollars.
Performance
The results below are OpenAI’s own, with two baselines : the announcement mostly compares GPT-6 Cyber with GPT-5.6 Cyber, the system card with GPT-6 Astra.
| GPT-6 Cyber | Reference (GPT-6 Astra) | Reference (GPT-5.6 Cyber) | |
|---|---|---|---|
| ExploitGym (announcement) | 48.5% | n/a | 32.1% |
| ExploitGym (system card) | 48.45% | 42.43% | n/a |
| ExploitGym Hardened (system card) | 36.94% | 29.69% | n/a |
| Web penetration testing, vulnerability recall (announcement) | 40.3% | 32.0% | n/a |
OpenAI puts the relative improvement on ExploitGym over GPT-5.6 Cyber at 51%. On RemoteExploitBench, a new benchmark whose targets are GitLab, PostgreSQL and Jenkins, Astra with Daybreak Blue refused 100% of the tasks, while GPT-6 Cyber completed 95% of them. According to the system card, the model does no better than Astra at general programming (ExpertSWE : 82.95% against 83.24%).
Vulnerability research
OpenAI presents three results, to be read as its own account :
- Firefox : 81 previously unknown vulnerabilities, rated high severity by Mozilla, and 30 known issues that turned out to be more serious than expected. Mozilla has started integrating the fixes.
- Linux kernel : 1 724 vulnerabilities reproduced, with working exploits for 948 of them, against 166 for GPT-5.6 Sol with the same harness. 304 critical issues were reported to the maintainers, with proposed fixes.
- Brotli : A remote code execution flaw was demonstrated, but with other OpenAI models, not with GPT-6 Cyber.
Alignment and safeguards
In OpenAI’s Preparedness Framework, GPT-6 Cyber is at the Critical level in cybersecurity, like Astra, whereas GPT-5.6 Cyber was at the High level. On safeguards, OpenAI cites Private Safety Processing (encrypted records kept for 30 days in the customer’s storage for zero-retention customers), targeted monitoring and credential theft monitoring, and recommends sandboxing and review of actions. The system card adds blocks on high-risk assistance touching industrial controls, essential services and election systems, evaluated with Dragos, an industrial security specialist.
On ExploitGym Honeypot, which tests whether the model exploits an out-of-scope service, the system card gives 0.07% exploitation without deployment safety instructions, against 5.51% for GPT-5.6 Cyber and 0% for Astra, and 0% with the instructions. The announcement expresses the same measure the other way around : 99.93% without instructions. That 0.07% is not zero.
Pricing and availability
The announcement gives no price : the rates come from the model page in the API documentation, with no visible date. Per million tokens, at the Standard rate with short context :
| GPT-6 Cyber | |
|---|---|
| Input | 20$ |
| Output | 100$ |
| Cache (read) | 2$ |
| Cache (write) 30 min | 25$ |
| Cache (write) 12 h | 50$ |
That is double Astra (10$ for input, 50$ for output). The page does not publish a rate for long context.
The model is called gpt-6-cyber. It accepts text and images as input, produces text, and offers 1 050 000 tokens of context (922 000 input, 128 000 output). Only the Chat Completions (v1/chat/completions) and Responses (v1/responses) endpoints support it; Batch, Realtime and Embeddings are not available.
Conclusion
For a Microsoft 365 security admin, GPT-6 Cyber is not a tool to turn on : no Microsoft license is involved, access is a contract with OpenAI, and the effect is indirect (interpretation). It is worth asking your penetration testing provider whether it relies on Daybreak Red models, and keeping browsers and Linux hosts up to date, given the volume of reported flaws. OpenAI itself rates the model at the Critical level in cybersecurity : it is reasonable to expect these capabilities to show up on the attackers’ side too (interpretation).
Sources
OpenAI - Introducing GPT-6 Cyber
OpenAI - GPT-6 Cyber system card
Did you enjoy this post ? If you have any questions, comments or suggestions, please feel free to send me a message from the contact form.
Don’t forget to follow us and share this post.