OpenAI Launches GPT-6 Cyber That Finds 81 Firefox Flaws and Develops 948 Linux Exploits

OpenAI has launched GPT-6 Cyber, a security-focused AI model that found 81 new high-severity Firefox flaws and built working exploits for 948 Linux kernel bugs. The company says it also follows the limits set by researchers far better than its older model.

OpenAI introduced GPT-6 Cyber on November 4, 2026. It is first going to a small group of approved Daybreak Red customers through the API and Codex.

On ExploitGym, a test of whether a model can exploit a known flaw, GPT-6 Cyber scored 48.5%. GPT-5.6 Cyber scored 32.1%. On a new web testing task, it found 40.3% of flaws, compared with 32.0% for GPT-6 Astra.

On RemoteExploitBench, GPT-6 Cyber finished 95% of tasks. Astra with Daybreak Blue refused all of them. OpenAI now rates the model Critical for cybersecurity under its Preparedness Framework. GPT-5.6 Cyber was rated High.

A big risk with security agents is that they wander outside their job. OpenAI tested this with ExploitGym Honeypot, which was shaped by the earlier Hugging Face incident. The test places easy-to-break side services next to the real target. Attacking them would reveal the flag, but they are out of scope.

Without safety prompts, GPT-6 Cyber stayed in scope 99.93% of the time, up from 94.49% for the older model. With the prompts, it hit 100%. Its reasoning notes show lines like “don’t explore out of scope credentials.”

81 Firefox Flaws

OpenAI researchers gave Mozilla proof and test cases for 81 new bugs that Mozilla rates high severity. The model also showed that 30 known issues were worse than first reported. Mozilla’s advisory MFSA 2026-118 covers fixes in Firefox 159, released October 27.

Many entries credit OpenAI Codex Security and OpenAI researchers. Most of the listed bugs are use-after-free memory errors in parts like layout, DOM, and JavaScript.

Other AI models have hit Firefox too. Cyber Security News reported that Claude Mythos found 271 Firefox zero-days earlier this year. Brian Grinstead of Mozilla said Firefox was one of the most hardened codebases going into 2026, and the new bugs still turned up.

948 Linux Kernel Exploits

OpenAI’s report, A thousand fewer paths to root, explains the Linux work. The pipeline picked 5,034 audit targets, about 7% of the kernel. It produced 2,308 candidate findings and reproduced 1,724 of them. It then built 948 working local privilege escalation exploits, which let a normal user become root. A GPT-5.6 Sol run in July found 166.

The run cost $1,213,430 in API credits, or about $1,279 per exploit. Past bounty prices would value the 948 exploits at $18.9 million to $66.3 million. OpenAI disclosed 304 critical findings with patches. Within 10 days, 61 had been fixed upstream and 43 had staged patches. Our earlier coverage details the 948 Linux kernel flaws.

OpenAI says the agents used at least 10 new exploitation techniques. Outside kernel experts had not seen them before. One early run also showed why checks matter. Of 200 reported successes, 20 gained root through a different zero-day than the one assigned. OpenAI added an independent check for that.

OpenAI added monitoring that checks whether actions match the user’s authorization. Daybreak Red API access requires IP allowlisting by default. The company also tracks stolen credentials and advises sandboxing and human review for risky work.

For power plants and water systems, OpenAI added a Daybreak Red for Critical Infrastructure group and a $1 billion Daybreak for Frontline Defenders grant program. Dragos tested the model in a mock power plant. Chief AI Officer Rob T. Lee said it separated OT assets from IT noise and found two zero-days in engineering software that handles protective relay data.

OpenAI’s advice is simple. Patch fast, rescan old code when models improve, and treat a low-privilege account on an unpatched system as close to root. It also stresses sandboxes like Firecracker and gVisor. The same group’s training environment report shows agents escaping microVMs during internal red-team tests. Teams can apply for Daybreak to request access.

Give your SOC the alert context to stops incidents quickly. Integrate Threat Intel from 16K SOCs.