From OpenAI · · 2m
GPT-6 Cyber: Security Model for Vulnerability Research
OpenAI's GPT-6 Cyber finds and fixes vulnerabilities with 51% better security performance than GPT-5.6 Cyber.
In brief
OpenAI's GPT-6 Cyber finds and fixes vulnerabilities with 51% better security performance than GPT-5.6 Cyber. OpenAI's GPT-6 Cyber identifies 81 high-severity Firefox vulnerabilities and 948 Linux exploits, improving security research by 51% on ExploitGym benchmarks. Originally reported by OpenAI.
GPT-6 Cyber Launch
OpenAI introduces GPT-6 Cyber, a cybersecurity model trained to find and validate exploitable vulnerabilities while staying within researchers' scope.
Security Performance
GPT-6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, with 51% relative improvement.
Real-World Impact
GPT-6 Cyber uncovered 81 high-severity Firefox vulnerabilities and developed exploits for 948 Linux kernel vulnerabilities.
Security Teams Adoption
Cisco, Calif, and CodeRabbit use GPT-6 Cyber for vulnerability research and code review, with Mozilla releasing fixes for identified Firefox vulnerabilities.
Critical Cybersecurity Rating
GPT-6 Cyber classified as Critical in cybersecurity under OpenAI's Preparedness Framework, up from High for GPT-5.6 Cyber.
Alignment Improvements
GPT-6 Cyber avoids exploiting out-of-scope services in 99.93% of tests without deployment prompts, 100% with prompts.
Chain-of-Thought Reasoning
GPT-6 Cyber uses chain-of-thought to reason about staying within bounds: 'Can't assume extra API or attack controller not target'.
“Can't assume extra API or attack controller not target. We should only target binary/server.”
Real-World Testimonials
Firefox, Cisco, Palo Alto Networks, and CrowdStrike report GPT-6 Cyber improves vulnerability detection and exploit development.
Critical Infrastructure Focus
OpenAI partners with Dragos to test GPT-6 Cyber on critical infrastructure, finding zero-days in engineering software.
Safeguards for Security
GPT-6 Cyber includes protections against misuse, including private safety processing and IP allowlisting for critical systems.
Daybreak for Frontline Defenders
OpenAI's $1 billion Daybreak for Frontline Defenders grant subsidizes access for government and essential service providers.
Expanding Access
GPT-6 Cyber available through API and Codex to Daybreak Red customers, with broader access planned for all users.
Real-World Results
GPT-6 Cyber found 81 novel Firefox vulnerabilities and developed exploits for 948 Linux kernel vulnerabilities.
Critical Infrastructure Testing
Dragos tested GPT-6 Cyber on a mock power plant, finding two zero-days in engineering software.
Safeguards for Critical Systems
OpenAI refines safeguards to block harmful assistance without restricting authorized testing of critical infrastructure.
Next Steps
OpenAI expands partnerships with critical infrastructure providers and security teams to refine safeguards and detect threats.
The short version
OpenAI's GPT-6 Cyber identifies 81 high-severity Firefox vulnerabilities and 948 Linux exploits, improving security research by 51% on ExploitGym benchmarks.
