OpenAI Releases GPT-6 Cyber With "Critical" Rating, Access Limited to Select Daybreak Red Customers

OpenAI has unveiled GPT-6 Cyber, a model trained to find software vulnerabilities, build working exploits and propose patches. The company rates it "Critical" in the cybersecurity category of its Preparedness Framework, a step above its predecessor GPT-5.6 Cyber, which was rated "High". For now, only selected, verified customers get access.

What the model is

GPT-6 Cyber is a variant of GPT-6 Astra designed for security tasks. According to gbhackers.com, it was trained with reinforcement learning on activities such as reproducing crashes, building exploits and searching codebases. The goal is to find and confirm vulnerabilities while staying within the scope set by the researcher.

The "Critical" cyber rating is the same as the one given to the base model, GPT-6 Astra, The Decoder reports. According to gbhackers.com, the level is up from the "High" assigned to GPT-5.6 Cyber.

Benchmark results

On the ExploitGym benchmark, GPT-6 Cyber scored 48.5 percent and GPT-5.6 Cyber 32.1 percent. GPT-6 Astra in the Daybreak Blue configuration reached 42.4 percent, and 0.0 percent with safeguards enabled. On a new test of web apps without access to source code, the model scored 40.3 percent against 32.0 percent for GPT-6 Astra.

On RemoteExploitBench, which covers known vulnerabilities in GitLab, PostgreSQL and Jenkins, the model completed 95 percent of tasks. At the same time, gbhackers.com reports, neither GPT-6 Cyber nor Astra solved any of the 15 hardest FrontierCyber challenges. Irregular observed no successful attacks on fully hardened targets.

Firefox and the Linux kernel

In Firefox, the model found 81 previously unknown severe vulnerabilities. Mozilla accepted the reports and, according to OpenAI, has already shipped fixes. In the Linux kernel, the model reproduced 1,724 vulnerabilities and built working exploits for 948, and 304 critical findings were passed to maintainers with proposed patches.

We covered this experiment earlier, with compute costs of about $1.2 million. Now OpenAI ties it to the launch of the model itself. The Decoder notes that the claim about techniques unknown to experts depends on how "own approaches" is defined, and that researchers disagree about whether models really discover anything new.

Who gets access

The model is available through the API and Codex to a select group of Daybreak Red customers, with gradual expansion planned. According to gbhackers.com, organizations must pass identity verification, single sign-on and multi-factor authentication, while individuals must use FIDO2 keys. IP allowlisting is applied by default for Daybreak Red API access.

A Daybreak Red for Critical Infrastructure group is also being created for authorized testing of systems such as power plants and water treatment facilities, along with a Daybreak for Frontline Defenders program backed by a billion-dollar fund. OpenAI recommends running the model in isolated environments with human oversight and acknowledges that abuse attempts are rising.

Caveats

The Decoder points out that Epoch AI researchers showed the open GLM model could have exploited known, unpatched Firefox vulnerabilities. That undercuts the assumption that access restrictions alone are enough. For Polish security teams, it means offensive tools of this class will be limited to verified customers, but similar capabilities may appear outside such programs.

The full results are to be included in the system card published by OpenAI. Until independent tests appear, the launch numbers are best treated as the vendor's claims.