OpenAI released GPT-6 Cyber for cybersecurity research. It finds and fixes vulnerabilities better than GPT-5.6 Cyber. It scores 48.5% on ExploitGym versus 32.1% before. It stays within scope 99.93% of the time without prompts. With prompts it never strays. It uncovered 81 new Firefox flaws and 948 Linux exploits. Mozilla has already patched many Firefox issues. It also excels at web penetration testing. Daybreak Red enables deeper exploit work. Daybreak Blue supports validated vulnerability checks. OpenAI used it to harden its own training infrastructure. It is classified as Critical under OpenAI’s Preparedness Framework.
What changed
GPT-5.6 Cyber had lower exploit success and weaker scope adherence.
What it unlocks
Finding and validating exploitable vulnerabilities while staying within authorized scope.
- 48.5% ExploitGym score
- 32.1% prior model ExploitGym score
- 99.93% scope compliance without prompt
- 100% scope compliance with prompt
What you need to act on it
- Daybreak access
- verified defender status
Sources