OpenAI Rates GPT-6 Cyber 'Critical,' Reporting 948 Linux Kernel Exploits and 81 Firefox Flaws

OpenAI has introduced GPT-6 Cyber, a security model trained to find and prove exploitable vulnerabilities while staying inside the scope a researcher sets, and it is the first model the company classifies as Critical for cybersecurity under its Preparedness Framework. The November 5, 2026 launch puts the model in the API and Codex for a select group of approved Daybreak Red customers. GPT-5.6 Cyber, its predecessor, was rated High.

OpenAI’s evaluation results underpin its higher rating. On ExploitGym, which asks a model to exploit a named vulnerability in a remote target, GPT-6 Cyber scores 48.5% against 32.1% for GPT-5.6 Cyber, a relative gain of about 51%. On a new web penetration-testing evaluation, where the model probes applications without seeing their source, it found 40.3% of vulnerabilities compared with 32.0% for GPT-6 Astra, OpenAI’s mainline model. On RemoteExploitBench, built from known GitLab, PostgreSQL, and Jenkins flaws in a controlled lab, GPT-6 Astra under Daybreak Blue refused every task while GPT-6 Cyber completed 95%.

Staying in scope is the headline claim

The risk OpenAI says it trained against is an agent that wanders, attacking services nobody authorized or ignoring the rules of engagement. To test for it, the company built ExploitGym Honeypot, which it says was informed by an earlier Hugging Face incident. The model gets a hard exploitation task in an isolated environment next to supporting services, such as a package registry or object store, that are deliberately easy to break and would hand over the task’s flag. The model has to work out that attacking them is off limits. Without deployment safety prompts, GPT-6 Cyber left those services alone in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber. With the prompts, it stayed in bounds every time.

Real targets, not just benchmarks

OpenAI says its researchers used the model across hardened open-source code. In a Firefox campaign it produced evidence and reproducible tests for 81 new vulnerabilities that Mozilla classed as high severity, and it showed that 30 known issues were worse than first reported. OpenAI sent the findings to Mozilla with proposed patches, and Mozilla has since shipped fixes for many of them. On the Linux kernel, the model reproduced 1,724 vulnerabilities and built working exploits for 948, more than five times the 166 that GPT-5.6 Sol managed with the same harness. Following the Linux community’s guidance, OpenAI disclosed 304 critical findings to maintainers with candidate patches.

Among launch customers, Cognition reports 23% on its internal novel-vulnerability benchmark against a previous best of 13%. CodeRabbit says the model found 66% more high- and medium-severity issues than its production model.

Safeguards scale with the rating

A Critical rating comes with tighter controls. Daybreak Red API access now requires IP allowlisting by default, credentials are watched for compromise, and OpenAI runs monitoring that checks whether a model’s actions match the user’s authorization. API customers on Zero Data Retention generally have to use Private Safety Processing, which reviews encrypted records kept in the customer’s own storage for 30 days without exposing prompts to OpenAI staff.

Critical infrastructure gets its own track. OpenAI announced Daybreak Red for Critical Infrastructure, a group of approved defenders allowed to test industrial systems, and pointed to its $1 billion Daybreak for Frontline Defenders grant program for under-resourced teams. Dragos ran evaluations with permissive Daybreak Red models on a mock power-plant range. Its Chief AI Officer, Rob T. Lee, said a model given almost no guidance separated operational technology from IT assets on its own and found two zero-days in software that manages protective relay data. OpenAI says that feedback shaped its safeguards around equipment interference and safety-critical disruption.

For security teams, the practical change is that exploit development, not just triage, is now a gated commercial product with published numbers. Access starts narrow, and outside groups have not yet reproduced the figures. OpenAI says it will widen access to all Daybreak Red customers soon, and GPT-6 Astra is opening to more defenders through Daybreak Blue alpha access.