cd /news/ai-safety/gpt-6-cyber-948-linux-exploits-81-fi… · home › topics › ai-safety › article
[ARTICLE · art-171136] src=byteiota.com ↗ pub= topic=ai-safety verified=true sentiment=· neutral

GPT-6 Cyber: 948 Linux Exploits, 81 Firefox Flaws

OpenAI released GPT-6 Cyber on November 5, a model its Preparedness Framework rates Critical, after running it on the Linux kernel and finding 948 working local privilege escalation exploits plus 81 previously unknown high-severity Firefox vulnerabilities, at a total API cost of $1.2 million. OpenAI reported the Linux research on November 4 in "A Thousand Fewer Paths to Root," where the pipeline narrowed the kernel's 5-million-plus lines to 5,034 audit targets and, from 2,308 candidate findings in a mid-October 2026 run, dynamically reproduced 1,724 and generated proofs of concept for 291 information-leak vulnerabilities and 948 LPE exploits; Mozilla has patched most of the Firefox flaws. Access is restricted to vetted organizations holding Daybreak Red approval, priced at $20 per million input tokens and $100 per million output tokens, with the full Linux kernel run costing $1,213,430, or about $1,279 per working LPE exploit.

read4 min views2 publishedNov 17, 2026
GPT-6 Cyber: 948 Linux Exploits, 81 Firefox Flaws
Image: Byteiota (auto-discovered)

OpenAI shipped GPT-6 Cyber on November 5. Before the press release went out, they had already run it on the Linux kernel. The pipeline found 948 working local privilege escalation exploits and identified 81 previously unknown high-severity vulnerabilities in Firefox. The total API tab: $1.2 million. OpenAI rates the model Critical — its highest cybersecurity tier — and is restricting access to vetted organizations only. Whether that restraint holds is the more interesting question.

948 Linux Exploits: How the Pipeline Works #

The Linux research was published November 4 as A Thousand Fewer Paths to Root. The pipeline runs in stages. Scout agents narrow the Linux kernel’s 5-million-plus lines down to 5,034 audit targets — a 93% reduction. Analyst agents then run vulnerability scans on those targets. From there: dynamic validation, exploitability assessment with proof-of-concept generation, and patch drafts. Human review happens before anything reaches kernel maintainers.

The mid-October 2026 run produced 2,308 candidate findings. Of those, 1,724 were dynamically reproduced. The pipeline generated proofs of concept for 291 information-leak vulnerabilities and 948 working LPE exploits for a commonly used Linux distribution. One early run caught a noteworthy QA failure: 20 of 200 reported successes had achieved root through a different zero-day than assigned. Human review caught it. The pipeline is not autonomous in the deployment sense — it is autonomous in the discovery sense.

The Firefox work followed a similar pattern. GPT-6 Cyber identified 81 high-severity zero-days that prior frontier models had missed. Mozilla has released patches for the majority of them.

The Alignment Numbers Actually Matter Here #

GPT-6 Cyber is better at staying inside its authorized scope than its predecessor by a meaningful margin. Without safety prompts, it remained in scope 99.93% of the time — up from 94.49% for GPT-5.6 Cyber. With deployment safety prompts active, it hit 100% in OpenAI’s honeypot tests.

That 5.5-percentage-point gap is not academic. At scale, a 5.5% out-of-scope rate on a model that generates hundreds of exploits per run is a significant liability. 0.07% is still not zero, but it is a materially different risk profile. OpenAI’s Preparedness Framework rates GPT-6 Cyber as Critical — the highest tier, reserved for models capable of finding and exploiting hardened systems without human guidance at each step. It earned that rating. The access controls are built around the assumption that the model is genuinely dangerous without them.

On benchmarks: GPT-6 Cyber scores 48.5% on ExploitGym — a 51% relative improvement over GPT-5.6 Cyber’s 32.1%. It completed 95% of tasks in OpenAI’s RemoteExploitBench evaluation. These are not marginal gains.

Metric GPT-5.6 Cyber GPT-6 Cyber
ExploitGym (pass@1) 32.1% 48.5%
In-scope (no safety prompt) 94.49% 99.93%
Linux LPEs (same harness) 166 948
Preparedness tier High Critical

Who Gets Access and What It Costs #

Access requires being an organization (individuals cannot apply), holding existing Daybreak Red approval, and then separately requesting the gpt-6-cyber model. Projects must have Daybreak Red explicitly enabled, with hardware security keys and allowlisted outbound IPs required. OpenAI is rolling out first to a select group of approved Daybreak Red customers.

Pricing: $20 per million input tokens, $100 per million output tokens — roughly twice GPT-6 Astra’s rate. The full Linux kernel run cost $1,213,430, or about $1,279 per working LPE exploit.

That figure deserves scrutiny. A traditional red team engagement that surfaces and validates hundreds of kernel-level privilege escalation vulnerabilities would cost significantly more and take far longer. From a defensive budget standpoint, $1,279 per patched attack vector is a credible value proposition. From an offensive one, the math is identical. The practical constraint is Daybreak Red vetting — in August, TechCrunch reported some researchers had access revoked without explanation. The gatekeeping is real, and it is not purely technical.

The Patch the Planet Connection #

GPT-6 Cyber feeds into Patch the Planet, a Daybreak initiative with Trail of Bits and HackerOne that runs AI-assisted vulnerability discovery against open-source projects and routes validated patches to maintainers. The human review layer is intentional — maintainer burnout from unfiltered AI-generated bug reports is a known problem, and the pipeline is designed to deliver patches rather than raw findings.

Week one (using GPT-5.5-Cyber) covered 19 projects, produced 64 pull requests, and filed 51 issues. GPT-6 Cyber is now in the pipeline. Those numbers will move.

The model is not publicly available, and OpenAI has been deliberate about that. The case for releasing it at all rests on proactive defense: finding and patching vulnerabilities before attackers do is worth enabling a small number of vetted organizations to use a Critical-tier model. That argument holds as long as the vetting does. The 0.07% out-of-scope rate and the August access revocations are both worth watching as the program scales.

── more in #ai-safety 4 stories · sorted by recency
── more on @openai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gpt-6-cyber-948-linu…] indexed:0 read:4min 2026-11-17 · —