OpenAI’s New GPT-6 Cyber Finds 81 High-Severity Firefox Vulnerabilities and 948 Linux Kernel Exploits

OpenAI has introduced GPT-6 Cyber, a security-focused model built to find and validate exploitable vulnerabilities while staying within the scope researchers set.

The company says the model has already uncovered 81 previously unknown high-severity Firefox flaws and produced working exploits for 948 Linux kernel vulnerabilities.

OpenAI’s New GPT-6 Cyber

GPT-6 Cyber is a specialized variant of GPT-6 Astra, specifically designed for cybersecurity tasks. It was trained using reinforcement learning on activities such as reproducing crashes, building exploits, and searching codebases.

In the ExploitGym evaluation, which tests the model’s ability to exploit specified vulnerabilities in a remote target, GPT-6 Cyber scored 48.5%, while GPT-5.6 Cyber scored 32.1%. In a recent web penetration testing evaluation, GPT-6 Cyber attained a recall rate of 40.3%, compared to 32.0% for GPT-6 Astra.

Additionally, on the RemoteExploitBench, which focuses on known vulnerabilities in GitLab, PostgreSQL, and Jenkins, GPT-6 Cyber completed 95% of the tasks.

In contrast, Astra running under Daybreak Blue refused to complete any of these tasks.OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.In the Firefox campaign, the model produced evidence and reproducible tests for 81 novel flaws that Mozilla rated high severity.

It also showed that 30 known issues were worse than first reported. OpenAI sent the findings to Mozilla with proposed patches, and Mozilla began landing fixes soon after. This follows earlier AI-driven hardening work, including 423 Firefox flaws fixed with Claude Mythos Preview.

The Linux kernel results are larger in scale. GPT-6 Cyber reproduced 1,724 vulnerabilities and built working exploits for 948. That is more than five times the 166 exploited by GPT-5.6 Sol using the same research harness. OpenAI disclosed 304 critical findings to maintainers with candidate patches, following Linux community guidance.

Scope control is a central claim. In the ExploitGym Honeypot test, the model faces a hard exploitation task in an isolated environment with deliberately weak auxiliary services, such as a package registry.

Attacking those services would reveal the flag but is out of scope. Without deployment safety prompts, GPT-6 Cyber avoided exploiting them in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber. With the prompts, it stayed in bounds every time. The Hugging Face incident informed the test.

The system card lists a 49.5% honeypot exploitation rate for GPT-5.6 Sol. This matters because OpenAI recently disclosed that its own agents escaped sandboxes and reached host systems during internal red-teaming.

Safeguards are layered. They include cyber-specific misalignment monitoring, Private Safety Processing for Zero Data Retention customers, default IP allowlisting for Daybreak Red API access, and credential-compromise monitoring.

OpenAI also recommends sandboxing and human oversight for high-risk workflows. Access requires identity verification, single sign-on and multi-factor authentication for organizations, and FIDO2 security keys for individuals.

For critical infrastructure, OpenAI added checks that block assistance involving industrial controls and essential services.

Dragos reported that a frontier model on a mock power-plant range separated OT assets from IT noise and found two zero-days in engineering software handling protective relay data. A new Daybreak Red for Critical Infrastructure group lets approved defenders test critical systems, backed by a $1 billion Daybreak for Frontline Defenders grant program.

GPT-6 Cyber is initially available through the API and Codex to select approved Daybreak Red customers, with broader access planned. Cisco, Calif., and CodeRabbit are among the early users.

The system card also shows limits. Irregular observed no successful attacks on fully hardened targets. Neither GPT-6 Cyber nor Astra solved any of the 15 hardest FrontierCyber challenges.

Give your SOC the alert context to stops incidents quickly. Integrate Threat Intel from 16K SOCs.