OpenAI Expands Daybreak With GPT-5.5-Cyber to Help Defenders Patch Security Flaws

What happened

OpenAI released the full GPT-5.5-Cyber to trusted defenders under its Daybreak program and unveiled the Patch the Planet initiative to secure the open-source ecosystem.

Context and impact

GPT-5.5-Cyber set SOTA on three cyber benchmarks (CyberGym 85.6%, ExploitGym 39.5%, SEC-bench Pro 69.8%). It already identified security components across 30M+ lines of Linux Kernel code and generated 8 kernel info-leak PoCs and 24 LPE exploits. The move responds to Anthropic's Mythos work and kicks off a 'frontier defender' model race.

Details

• Full version restricted to the Trusted Access for Cyber program
• Benchmarks: CyberGym 85.6% (+3.8), ExploitGym 39.5% (+13.55), SEC-bench Pro 69.8% (+6.7)
• Partners: Trail of Bits, HackerOne
• 30+ open-source projects on board (cURL, Go, Python, Sigstore, pyca/cryptography, aiohttp, NATS, freenginx)
• Audit covered 30M+ LOC of the Linux Kernel