Zero Day Room
Live
Vulnerabilities

GPT-6 Cyber Discovers Critical Firefox and Linux Vulnerabilities

OpenAI's GPT-6 Cyber model identified 81 previously unknown high-severity Firefox vulnerabilities and 948 distinct Linux kernel exploits.

OpenAI's GPT-6 Cyber model identified 81 previously unknown high-severity Firefox vulnerabilities and 948 distinct Linux...

OpenAI's GPT-6 Cyber discovered 81 previously unknown high-severity Firefox vulnerabilities and 2,308 potential Linux kernel flaws. The model developed working local privilege escalation exploits for 948 distinct Linux kernel vulnerabilities, which could allow an unprivileged user to gain root access. Mozilla has started releasing fixes for the Firefox issues, and the Linux kernel exploit findings are being sent to maintainers with proposed patches for review.

OpenAI unveiled GPT-6 Cyber, a version of GPT-6 Astra trained to find software vulnerabilities and test exploit methods. The model is classified as having "Critical" cybersecurity capabilities under OpenAI's Preparedness Framework, an increase from the "High" classification given to its predecessor. It is initially available through API and Codex to selected, approved Daybreak Red customers, with plans to extend access to all customers in that program over the coming weeks. The model is not publicly available.

Model Classification and Access

GPT-6 Cyber is OpenAI's most capable and best-aligned cybersecurity model. It was trained to stay within authorized limits, respect restrictions, and report results accurately. The model is designed for authorized security testing of websites and applications, helping to confirm exploitability, look for other attack routes, strengthen patches, and assist with deployment. OpenAI will expand GPT-6 Cyber access to all Daybreak Red customers in the coming weeks.

Validation and Impact

Researchers used GPT-6 Cyber alongside other OpenAI models to validate exploit potential. They demonstrated how a critical Brotli vulnerability could allow remote code execution. Mozilla has released fixes for many of the Firefox flaws identified. Brian Grinstead, a Distinguished Engineer at Firefox, stated that GPT-6 Cyber is a world-class model for cyberdefense. He noted that even after significant hardening, the model uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped ship fixes to millions of users. The goal is security working at 'agent speed'.

Performance and Safety

GPT-6 Cyber significantly outperforms earlier AI models in finding software vulnerabilities and validating their impact. In benchmark testing, it demonstrated superior capability in authorized security testing.

ModelExploitGym Benchmark Score
GPT-6 Cyber48.5%
GPT-5.6 Cyber32.1%

In safety evaluations, GPT-6 Cyber avoided exploiting an out-of-scope service in 99.93% of tests without deployment safety instructions and 100% of the time with them. OpenAI used the model to help secure its own systems, including AI training infrastructure, where it identified vulnerabilities that allowed agents to escape isolated environments.

Security Safeguards and Partnerships

OpenAI implemented multiple new or strengthened safeguards for this model. These include monitoring for unauthorized actions, targeted threat monitoring, Private Safety Processing, credential protection, and checks to block high-risk cyber assistance. Sandboxing, action review, and human oversight were recommended for high-risk workflows.

The company tested protections with industrial cybersecurity firm Dragos. Using cyber-permissive Daybreak Red models in a controlled lab, they covered scenarios from network discovery to unauthorized device actions. Dragos reported that GPT-6 Cyber found two zero-days in engineering software managing protective relay data, exposing protection settings and breaker trip logic on a mock power plant range.

OpenAI plans to expand partnerships with security providers, critical infrastructure operators, and national laboratories to evaluate AI-assisted testing of industrial systems. The company has increased investment in threat investigations to detect model use involving operational technology and industrial control systems. It also worked with Cloudflare to deploy web application firewall rules that protect sites before a bug was announced.

OpenAI warned that defenders have a limited window to use these increasingly powerful AI models to discover and patch vulnerabilities before similar capabilities become more widely available. Security experts note that finding flaws is the easy part, but getting them fixed is harder. Mozilla's ongoing release of patches for the discovered Firefox vulnerabilities continues.

Related coverage

More from Vulnerabilities