OpenAI's GPT-6 Cyber Scores 48.45% on ExploitGym, but Only Vetted Daybreak Red Customers Get It

OpenAI released GPT-6 Cyber on Thursday, November 5, a security-tuned version of GPT-6 Astra that it rates "Critical" for cybersecurity under its Preparedness Framework. The previous model, GPT-5.6 Cyber, was rated "High." Access is limited to approved customers in the Daybreak Red program, through the API and Codex.

OpenAI's system card, published the same day, reports that GPT-6 Cyber scored 48.45% on ExploitGym, a test of whether an agent can turn a known vulnerability into a working exploit. GPT-6 Astra scored 42.43% on the same test. All figures here are OpenAI's own.

ExploitGym: 48.45% against 42.43% for Astra

The Hacker News and itdaily both led with a different comparison: 48.5% against 32.1% for GPT-5.6 Cyber. The system card's own comparison is against Astra, which narrows the gap to about six points.

1. GPT-5.6 Cyber32.1 %
2. GPT-6 Astra42.43 %
3. GPT-6 Cyber48.45 %

Source: OpenAI GPT-6 Cyber system card (Astra, Cyber); The Hacker News (GPT-5.6 Cyber), accessed 2026-11-09. Vendor-supplied.

On ExploitGym Hardened, the card gives 36.94% for GPT-6 Cyber and 29.69% for Astra. The card also measures how often a model exploits a service it was told to leave alone. Lower is better.

Model	Honeypot exploitation, no safety prompt
GPT-5.6 Sol	49.5%
GPT-5.6 Cyber	5.51%
GPT-6 Astra	0%
GPT-6 Cyber	0.07%

With deployment safety instructions added, GPT-6 Cyber's rate is 0%, according to the card.

Firefox, Linux and Brotli results

OpenAI says researchers used the model to find 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity, per Help Net Security. Mozilla has shipped fixes for many of them. Help Net Security quoted Brian Grinstead, Distinguished Engineer at Firefox: "GPT 6 Cyber is a world-class model for cyberdefense."

The model also produced working exploits for 948 Linux kernel vulnerabilities, according to The Hacker News and Help Net Security. itdaily reported that OpenAI has shared 304 of those findings with maintainers. The Hacker News listed a critical Brotli flaw, CVE-2026-111497, with a CVSS score of 9.2.

Rob T. Lee, Chief AI Officer at Dragos, said his company "put a frontier model on a mock power plant range and told it almost nothing," as quoted by Help Net Security. Neither the Firefox nor the Dragos statement is an independent test. Both come from partners.

Who can use it

The card says individuals in Daybreak Red must pass identity and liveness checks, show professional suitability, provide a legal attestation and use a physical FIDO2 key. Organizations need single sign-on, multi-factor authentication and managed devices. API IP allowlisting is the default.

OpenAI also blocks high-risk help involving industrial controls, essential services and election systems. Help Net Security said OpenAI plans to extend access to every customer in the program. OpenAI has not published a price.

What the card says it cannot do

The card says ExploitGym uses historical vulnerabilities, so contamination from training data may inflate scores. It states that "Reliable exploitation of these recently disclosed vulnerabilities remains challenging for GPT-6 Cyber," and gives 42% on one internal set.

Irregular, an outside assessor named in the card, observed no successful attacks on fully hardened targets. Neither Astra nor GPT-6 Cyber solved the 15 hardest FrontierCyber challenges. The card also says GPT-6 Cyber is slightly less likely than Astra to stop when warnings warrant it.

Earlier coverage on The Terminal looked at a Cloudflare security audit skill and at Anthropic's agent memory study, both bearing on how agent work is measured.

OpenAI has not named a date for wider access. The Hacker News reported separately that from November 29 OpenAI plans to use content from Daybreak Blue sessions on personal accounts to train safeguards, even with the improve-the-model setting off.