OpenAI releases GPT-6 Cyber to vetted defenders

OpenAI announced GPT-6 Cyber, which its system card describes as a cyber-specialised variant of GPT-6 Astra built for authorised security work: vulnerability discovery, exploit development and web penetration testing. It is not publicly available. Access runs through Daybreak Red, the higher-trust tier of the Daybreak programme, with single sign-on and multi-factor authentication for organisations and identity and liveness checks, legal attestation and a hardware security key for individuals.

The system card gives the model the same Preparedness designations as Astra: Critical for cybersecurity, High but not Critical for biology and chemistry, and below High for AI self-improvement. Astra had been the first model OpenAI flagged at the Critical cyber tier. The card lists 48.45% on ExploitGym against 42.43% for Astra; the launch post compares it with the earlier GPT-5.6 Cyber, at 48.5% against 32.1%. On a test of whether the model exploits services it was told were out of scope, OpenAI said it stayed in bounds 99.93% of the time without its deployment safety prompt, against 94.49% for GPT-5.6 Cyber, and every time with the prompt.

The card carries its own caveats. It warns that ExploitBench, whose vulnerabilities date back to 2023, “raises a risk of contamination that may inflate the reported results”, and says OpenAI built an internal evaluation to reduce that. Irregular, an outside evaluator, found performance broadly comparable to Astra’s: 90 of 288 FrontierCyber challenges against Astra’s 96, and neither model solved any of the 15 hardest. The card also names a collaboration with NIST’s CAISI on cyber measurement but reports no results from it.

OpenAI said the model produced evidence and reproducible tests for 81 previously unknown vulnerabilities that Mozilla classed as high severity in Firefox, and that Mozilla had shipped fixes for many of them. A report published the day before, “A thousand fewer paths to root”, described a pipeline run in mid-October with “an upcoming cyber model” that reproduced 1,724 Linux kernel vulnerabilities and built working local privilege-escalation exploits for 948, at an API cost of about $1.2 million; human researchers reviewed the output before reporting. The launch post identifies that model as GPT-6 Cyber, compares the 948 with 166 for GPT-5.6 Sol on the same harness, and says 304 critical findings went to the Linux maintainers with candidate patches.

Security vendors quoted by The Next Web were cautious. Noma Security’s Chris Hughes said that at this scale even a small out-of-scope fraction means actions against unauthorised systems, and noted that fewer than a quarter of flaws on OpenAI’s Patch the Planet dashboard had been fixed.

The release came a week after Anthropic described exploit chains built by an unreleased model and Google released Gemini 4 Carbon to cyber defenders. Amazon had made the earlier GPT-5.6 Cyber available through Bedrock in AWS GovCloud on 30 October, with OpenAI’s own post following on 4 November.