GPT-6 Cyber scores 48.5% on ExploitGym and is rated Critical; Daybreak Red access gated to select customers

OpenAI introduced GPT-6 Cyber on Nov 5 as a cybersecurity model trained to find and validate exploitable vulnerabilities while staying within the scope researchers set. On ExploitGym, which tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber — about 16 percentage points, which OpenAI frames as a 51% relative improvement. On a separate alignment eval, ExploitGym Honeypot, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt (up from 94.49% for GPT-5.6 Cyber) and 100% with that prompt. OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. On RemoteExploitBench, GPT-6 Astra with Daybreak Blue refused 100% of tasks; GPT-6 Cyber completed 95%; its only failure was an exploit it could not get to work remotely. On a web pentest eval, GPT-6 Cyber reached 40.3% recall versus 32.0% for GPT-6 Astra. OpenAI says its researchers used it to identify 81 previously unknown high-severity Firefox vulnerabilities, and Mozilla has released fixes for many of them; the earlier Linux privilege-escalation pipeline is already covered on this desk and is not today's headline. Access starts through the API and Codex for select approved Daybreak Red customers, with expansion planned soon; GPT-5.6 Cyber is being made available to all Daybreak Red customers, and Daybreak Blue alpha access for GPT-6 Astra is expanding. OpenAI links a system card for full results and evaluation settings. Why it matters: A model that both finds more working exploits and refuses out-of-scope ones more often changes how security teams plan around AI helpers: capability and restraint are supposed to move together. OpenAI's own evals put GPT-6 Cyber well ahead of GPT-5.6 Cyber on ExploitGym, near-perfect on the honeypot scope test, and rate it Critical. Most readers cannot call the model yet — Daybreak Red access is gated to approved customers — so the near-term move is to treat the release as a planning signal for red-team rules and disclosure workflows, not as a tool you can open today. The Linux 948-exploit pipeline was already on this desk on Nov 5; today's news is the model release, the Critical rating, and the gated Daybreak access.

Try this: Open one AI-assisted security workflow you already run and write down the scope boundaries you expect the model to honor (allowed hosts, forbidden services, human gate before out-of-scope tool calls). For each boundary, add one concrete check. Compare those checks with the behavior OpenAI says ExploitGym Honeypot measures — staying off easy out-of-scope services — before you plan any Daybreak Red application. Treat the ExploitGym and honeypot percentages as vendor evals until someone outside OpenAI reproduces them.