OpenAI demoed Daybreak Red Team in early access and Daybreak Defense Factory built on GPT6 Cyber, with field results in Firefox, PHP and Linux that cover remote code execution and local privilege escalation which is an attack that turns ordinary local access into full root control.
Defenders don't lack bugs, they lack proof of which ones will hurt.
Nancy Wang, who leads cyber applied AI at OpenAI, positioned Daybreak as continuous defense productized from internal work. Red Team, which is automated offensive testing that tries to break running software like an attacker would, starts when an engineer picks a target such as a staging environment and lets Codex draft a campaign from connected context, rate limits, credentials and allowed traffic. A fleet of agents then divides tasks through a shared queue and knowledge base while a supervisor agent checks for drift, a structure OpenAI describes as safety in depth.
Defense Factory, which is the closing loop that finds, validates and fixes code flaws, runs in reproducible cloud development environments where agents trace a flaw to the affected file and use code history and team structure to route it to the right owner. The output is a merge-ready pull request, which is a proposed code change packaged with tests for a human to review and accept, plus a test that proves the fix works. Wang said new commits get reviewed the same way, so fixes arrive as code ships rather than months later in a backlog sprint.
Clint Gibler, who leads security research at OpenAI, said scanning the same code with a newer model surfaces new flaws and upgrades old ones, with low severity denial of service issues becoming remote code execution which is an attacker running commands from across the network. In his numbers, GPT6 Cyber reproduced 81 previously unknown high severity Firefox flaws and reframed 30 known low or medium issues as high severity, while one Linux pipeline validated patches for almost 1,000 local privilege escalations, more than five times the GPT-5.6 Cyber total from three months earlier, plus 10 novel exploitation techniques. Fouad, who presented the model results for OpenAI, put GPT6 Cyber at 52% on Exploit Gym which tests proof of concept creation, against about 38% for Anthropic Mythos 52 cited for comparison, with the next model near double that and around 80% on PatchBench which tests finding and fixing without regressions, and 99.9% in-scope behavior without production safeguards and 100% with them according to the honeypot eval which plants an easy out-of-scope bug to tempt the agent.
Greg said the PHP-shell case, a flaw that lived in PHP for 13 years, took two months of joint patching, variant analysis, bypass testing and government calls and consumed all available bandwidth from OpenAI and its partners. That is the scaling problem, and OpenAI points to Patch the Planet, which according to its Daybreak page brings OpenAI, Trail of Bits, Calif and HackerOne together to validate findings, test patches and coordinate disclosure, plus a billion dollar Daybreak defense fund for frontline defenders cited on the broadcast. Watch whether AI-generated pull requests land cleanly and whether disclosure can handle ten PHP-scale bugs at once instead of one.
