OpenAI (Nov 5, 2026) first-party introduces GPT-6 Cyber, a cybersecurity-specialized model for finding/validating exploitable vulnerabilities while staying inside operator-set scope — soft title: Security Watch + models ship, not a general GPT-6 consumer GA. Vendor claims major gains vs GPT-5.6 Cyber (incl. ~51% relative ExploitGym improvement + “step function” scope-stay evals); already used on security-critical codebases including Linux and Firefox. Access: initially API + Codex for select approved Daybreak Red customers, expanding over time; also launches Daybreak Red for Critical Infrastructure for approved defenders’ authorized testing, with tightened OT/ICS misuse safeguards. Related: OpenAI Cyber at the Frontier livestream Nov 12, 10:00 AM PT. Elevates overnight X discovery @gdb only after OpenAI primary openai.com/index/introducing-gpt-6-cyber/. Distinct from Thu MCPhantom MCP supply-chain card (hold).
Security Watch lead + models — prefer-window frontier cyber-agent ship with Critical-Infrastructure Daybreak tier after Thu agent-supply-chain book closed without a NEW cyber-model card.
More ▾ Less ▴
OpenAI’s November 5, 2026 blog introduces GPT-6 Cyber, a cybersecurity-specialized model for finding and validating exploitable vulnerabilities while staying inside operator-set scope. Soft title: Security Watch + models ship — not a general GPT-6 consumer GA. Vendor claims ~51% relative ExploitGym improvement vs GPT-5.6 Cyber and a step-function gain on scope-stay evals; already used on security-critical codebases including Linux and Firefox. Access starts via API + Codex for select approved Daybreak Red customers, with Daybreak Red for Critical Infrastructure for approved defenders’ authorized OT/ICS testing and tightened misuse safeguards. Related: Cyber at the Frontier livestream Nov 12, 10:00 AM PT. Elevates overnight X @gdb discovery only after the OpenAI primary. Stamp: OpenAI blog Nov 5 ≈ 01:09–01:31 HKT Fri after @OpenAI/@gdb posts ~17:09–17:31 UTC Nov 5. Distinct from Thu LIVE MCPhantom (held). Editor spot-check: OpenAI URL HTTP 200 into Fri close.
- Daybreak Red expansion cadence
- Nov 12 Cyber at the Frontier livestream
- Independent defender write-ups
- Any Linux/Firefox maintainer follow-through beyond vendor colour
OpenAI first-party blog Confirmed Nov 5. Vendor eval/exploit counts and Critical Preparedness self-rating are OpenAI’s. Access gated. Distinct from Thu MCPhantom card.