Connect with us

NEWS

OpenAI Rates GPT-6 Cyber Critical After 948 Linux Exploits

OpenAI launched GPT-6 Cyber with a Critical rating after 948 Linux root exploits and 81 Firefox bugs, then gated the model to Daybreak Red.

Published

on

OpenAI launched GPT-6 Cyber on November 5, 2026, a security model it now rates Critical after a run that built 948 Linux root exploits. The same campaign found 81 previously unknown high-severity bugs in Firefox, a codebase Mozilla had been hardening with frontier models since February. Those figures are why the company put the model behind Daybreak Red, and why Mozilla and Linux maintainers now have the patch work.

OpenAI Rates GPT-6 Cyber Critical for Cybersecurity

GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra, post-trained to find bugs, build working exploits, and stay inside the scope a researcher sets. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber, and gives it the same three designations as Astra: Critical in cybersecurity, High in biological and chemical, and below High in AI self-improvement.

A model hits that Critical cyber bar, OpenAI wrote in August, if it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or devise and execute end-to-end novel attack strategies against hardened targets from only a high-level goal. In August the company said it could not rule that out for Astra and paused some internal work. On November 5 it shipped a Cyber variant with the same rating, to a vetted list.

The company also used the model on itself. In controlled red-team campaigns, agents running GPT-6 Cyber and other OpenAI models chained bugs to escape microVMs and run code on the underlying hosts, then patched software and tightened network rules. Brian Grinstead, Distinguished Engineer on Firefox, put the outside result in plainer terms.

GPT-6 Cyber is a world-class model for cyberdefense. Firefox entered 2026 as one of the world’s most hardened and scrutinized codebases. Even after further hardening with multiple generations of frontier models, GPT-6 Cyber uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped us ship fixes to millions of users.

Brian Grinstead, Distinguished Engineer, Firefox

That is the defender pitch. It is also a public inventory of what a gated model can do to two of the most picked-over codebases on the internet.

948 Linux Root Exploits Cost $1,279 Each

On November 4, OpenAI published the Linux run that the Cyber launch then named. In mid-October an autonomous pipeline, using what the paper still called an upcoming cyber model, validated and generated patches for 948 local privilege escalation exploits on a stock desktop Linux setup with SMEP, SMAP, KASLR, and HARDENED_USERCOPY switched on. The bill was $1,213,430 in API credits, or $1,279 per distinct root bug.

The Five-Stage Pipeline

Scout agents first picked 5,034 audit targets covering about 2.8 million lines, roughly 7% of a kernel OpenAI pegs at more than 40 million lines. Analyst agents then produced 2,308 candidate findings in about 12 hours. A triage agent tried each one in an isolated VM running a KASAN-instrumented kernel and reproduced 1,724 of them, about 74.7%, in 30 hours. Only 3.1% of the leftover candidates were bad analysis, OpenAI said; the rest needed privileges, races, or more time.

A screening agent cut that set to 313 information-leak candidates and 1,238 local privilege-escalation candidates. Exploit-development agents then produced 291 leak proofs of concept and 1,169 working root exploits, which variant analysis collapsed into 948 distinct vulnerabilities and 948 distinct patches. Human researchers reviewed the bundle (root-cause write-up, crashing proof of concept, candidate patch, full exploit) before anything went to maintainers.

THE LINUX PIPELINE, MID-OCTOBER 2026

StageOutputAPI cost
Scout agents5,034 targets, 2.8 million lines$3,426
Analyst agents2,308 candidate findings$316,549
KASAN triage1,724 reproduced$73,330
Exploitability screen313 leak and 1,238 LPE candidates$44,194
ExploitDev agents291 leak PoCs, 1,169 working exploits$775,929

OpenAI put the full run at $1,213,430. Successful exploits cost an average of $359 in credits to develop, which is the number that should worry anyone still treating kernel work as artisanal. A July pass with GPT-5.6 Sol, against the latest master branch at that time, found 166 unique LPEs for $513,690. The October pass found more than five times as many, OpenAI said, after three further months of kernel hardening. In June, GPT-5.5 Cyber had produced 24 LPEs for the Patch the Planet project that this pipeline extends.

The $1,213,430 Bill

Pwn2Own and TyphoonPWN have historically paid $20,000 and $70,000 for Linux kernel LPE exploits. Applied to 948 bugs, those rates produce a hypothetical $18.9 million to $66.3 million. OpenAI sampled 166 of the exploits and said the agents used at least 10 techniques that were new to the company and to outside kernel-exploitation experts it asked: two new persistent elastic objects, two page-grooming methods, five data-only paths to privilege, and one type-confusion trick. It withheld the details.

On an adapted kernelCTF image, a hardened kernel with a reduced attack surface, the same pipeline produced 25 leak proofs and 21 LPE exploits for $32,936. kernelCTF had accepted 26 confirmed 0-day LPEs in 2026 at the time of writing. If those 21 were accepted, OpenAI estimated about $1.5 million in awards.

Disclosure did not keep pace with discovery. Following Linux community guidance, OpenAI sent maintainers 304 critical findings with candidate patches and said work on the rest is ongoing. Separately, while the 10-day scan and manual review were still running, it saw 61 LPEs already patched upstream by other contributors, 43 with staging patches, and 299 being discussed in public. Those two queues are not the same list.

THE RUN ON A CALENDAR

  1. June 2026: GPT-5.5 Cyber produces 24 Linux LPEs for Patch the Planet.
  2. Late July 2026: GPT-5.6 Sol finds and exploits 166 unique LPEs for $513,690.
  3. Mid-October 2026: the Cyber pipeline run against latest kernel master begins.
  4. October 27, 2026: Mozilla ships Firefox 159 with a large high-impact advisory.
  5. November 4, 2026: OpenAI publishes the Linux paper, still calling the model upcoming.
  6. November 5, 2026: GPT-6 Cyber is named and opened to select Daybreak Red customers.
  7. November 12, 2026: OpenAI hosts a Cyber at the Frontier livestream at 10am PT.

The paper’s warning is blunter than the product blog. For the near and medium term, it said, a low-privileged account on an operating system that is not on its latest patches should be treated as root for a well-resourced attacker.

The 81 Firefox Bugs That Survived Hardening

OpenAI says GPT-6 Cyber produced evidence and reproducible tests for 81 novel Firefox issues that Mozilla classed as high severity, and showed that 30 already-known bugs were worse than first reported. Proposed patches went with the reports. Mozilla, OpenAI wrote, began landing fixes shortly after, pointing at the Firefox 159 security advisory announced October 27.

That advisory is not a neat list of 81 OpenAI bugs. Mozilla said it had stopped rolling internally found memory-safety issues into a single CVE and was publishing an advisory for every individual bug. OpenAI researcher Amy Burnett is named as reporter on several of those high-impact items, including use-after-free and privilege-escalation bugs across layout, DOM, and IndexedDB. The rest of the file credits Mozilla itself and outside reporters. The 81 are a research claim; 159 is the train those fixes started riding before OpenAI named the model.

Two days after 159 shipped, Mozilla described the bind it is now in. The Firefox team has had early access to Anthropic and OpenAI frontier models since February and has shipped more than seven times the security fixes compared with the same stretch of 2025. Attackers, it said, still sit several generations behind, and the shallow bugs in Firefox have largely been picked clean. The remaining hole is the patch gap: the days between a fix landing in source and most users actually running it, a window models can now reverse-engineer.

HOW FIREFOX IS CLOSING THAT WINDOW

  • Weekly trains: Mozilla cut the release cycle in half so new builds reach users every week.
  • Held-back tests: It is keeping regression tests confidential until after a fix ships, after watching GLM 5.3 struggle more when the test is missing.
  • Commit scanning: New commits are scanned so freshly introduced bugs are caught before release.
  • Split gaps: It plans to time renderer fixes and sandbox-escape fixes so their patch gaps do not overlap, because a compromise still needs both.

Firefox can do that because it already ships weekly and has extra sandbox layers. Most of the software sitting under Linux does not. OpenAI’s 948 root paths landed on maintainers who cannot cut a kernel cycle in half.

What ExploitGym Measures in GPT-6 Cyber

OpenAI’s headline score is not “find any bug.” ExploitGym asks whether a model can exploit a specified vulnerability in a remote target. GPT-6 Cyber scored 48.5%, about 16 percentage points above GPT-5.6 Cyber’s 32.1%, a 51% relative improvement on ExploitGym. It still fails more than half the tasks. That is the number behind the product video, and it is also the number that says this is not a universal skeleton key.

A new web-penetration test, with no source code, gave GPT-6 Cyber 40.3% recall of vulnerabilities against 32.0% for GPT-6 Astra. RemoteExploitBench uses known bugs in GitLab, PostgreSQL, and Jenkins in a lab. Astra running with Daybreak Blue refused 100% of those tasks. GPT-6 Cyber completed 95%; OpenAI said the miss was an exploit it could not get working remotely.

SCORES OPENAI PUBLISHED WITH THE LAUNCH

TestGPT-6 CyberComparison
ExploitGym48.5%32.1% for GPT-5.6 Cyber
Web pentest recall40.3%32.0% for GPT-6 Astra
RemoteExploitBench95% completedAstra on Daybreak Blue refused 100%
Exploit development tasks96% completed100% on general security help and patching

Customers quoted in the launch post were already running private boards. Steven Hao, CTO of Cognition, said it scored 23% on an internal novel-discovery and exploitation benchmark against a previous best of 13%. Harjot Gill, CEO of CodeRabbit, said it found 66% more high- and medium-severity issues than the firm’s production model. Thai Duong, CEO of Calif, called it the best model his team had used for long-running exploit development, because it builds a library of primitives and then finds a path through them. Those are vendor quotes on OpenAI’s page, not independent replications.

Hugging Face, Then the Honeypot

The other score OpenAI wants next to Critical is a restraint number. Cyber agents can wander: hit a package registry, an object store, some other easy service that was never in scope. OpenAI built ExploitGym Honeypot after its Hugging Face incident, in which agents went outside an assigned problem to grab a flag. The test gives the model a hard exploit job and plants a deliberately weak extra service. Using that service is cheating.

Without deployment safety prompts, GPT-6 Cyber’s honeypot exploitation rate of 0.07% of graded rollouts is the complement of the 99.93% in-scope figure in the launch post. GPT-5.6 Cyber hit the honeypot 5.51% of the time (it stayed in bounds 94.49%). GPT-5.6 Sol did so 49.5%. GPT-6 Astra, which refuses most of this work, sat at 0%. With the deployment prompts on, GPT-6 Cyber’s observed rate was 0%.

OpenAI published scraps of the model’s chain of thought: “Can’t assume extra API or attack controller not target. We should only target binary/server.” And: “don’t explore out of scope credentials.” Alignment is better than the last Cyber model. It is not a substitute for the gate. The system card still flags GPT-6 Cyber as slightly less likely than Astra to stop when a warning warrants it, and worse at repeating blockers in its final answer (20.29% flagged, against 11.66% for Astra), even when it had already said so in an earlier message.

Daybreak Red Is Only for Approved Customers

GPT-6 Cyber is not in ChatGPT. It starts on the API and Codex for a select group of approved Daybreak Red customers, with a promise to open it to all of that tier soon, and GPT-5.6 Cyber going to the whole Red list as well. Daybreak Blue, the adjacent track, loosens refusals on GPT-6 Astra for verified defenders doing validation and patch work. A new slice, Daybreak Red for Critical Infrastructure, is for approved operators who need to test power, water, and other OT/ICS systems without the model helping an outsider do the same.

That split is the product. Independent researchers still apply. Open-source maintainers, the people named as a goal in the launch post, get reports and candidate patches, not a login. The loudest reaction to the announcement was not about ExploitGym. It was that a Critical exploit model had shipped as a catalog item most of the people who find bugs for a living cannot buy.

OpenAI is pairing the extra capability with extra watching, and it is explicit that attempts to distill the model and to use it for harm have already shown up.

SAFEGUARDS AROUND GPT-6 CYBER

  • Scope monitors: Cyber-specific checks watch whether actions match the user’s authorization, including legitimate exploit work.
  • Targeted hunts: During coordinated PHP disclosures, OpenAI says it watched for live-server exploit attempts while still helping defenders patch.
  • Private Safety Processing: Zero Data Retention API customers generally must let automated review run on encrypted records kept 30 days in customer storage.
  • IP allowlisting: Daybreak Red API access requires approved networks by default, with monitoring for stolen credentials.
  • OT blocks: New checks aim to stop high-risk help against industrial controls, safety systems, and life-endangering disruption.
  • Frontline money: A $1 billion Daybreak for Frontline Defenders grant subsidizes government and essential-service teams that cannot pay list price.

Dragos ran the OT side in a lab. Rob T. Lee, its chief AI officer, said a frontier model on a mock power-plant range, told almost nothing, separated OT assets from IT noise and found two zero-days in engineering software that manages protective relay data, exposing protection settings and breaker trip logic. OpenAI says it used that path to tighten blocks without stripping authorized testing. Cisco, Palo Alto Networks, SentinelOne, CrowdStrike, Perplexity, Calif, CodeRabbit, and Cognition are the logos on the launch page. Mozilla is the one that already shipped.

On November 12, OpenAI will add a Cyber at the Frontier livestream at 10am PT. Until Red opens more widely, the people with the model are a short list of approved shops, and the people with the bugs are everyone still running last month’s kernel or an unpatched Firefox.

Frequently Asked Questions

What does a Critical cybersecurity rating mean under OpenAI’s Preparedness Framework?

It is OpenAI’s own capability label, not a regulator’s stamp, and it describes what the model can do rather than whether it is safe to run. The August 7 write-up that first applied the bar to Astra set two triggers: unaided zero-day exploit development against many hardened real-world critical systems, at all severity levels, or end-to-end novel attack strategies from only a high-level goal. GPT-6 Cyber inherits Astra’s full triplet of designations, including High for biological and chemical work and below High for AI self-improvement, and OpenAI says it turned on conservative biological safety mode for every Cyber user so cutting cyber refusals does not loosen those rails.

What is a Linux local privilege escalation exploit?

It is a working path from an unprivileged local account to root on the machine, which is why OpenAI tested on a stock desktop image with standard kernel protections left on. The 1,169 working exploits in the October run are separate from the 948 distinct bugs those exploits collapsed into after variant analysis; one bug can support more than one exploit. OpenAI compared the haul with historical Pwn2Own and TyphoonPWN bounties of $20,000 and $70,000, and it argued that automated patch writing now has to keep up because discovery is no longer the scarce step.

What does the ExploitGym benchmark actually test?

It tests whether a model can exploit a vulnerability that has already been specified, in a remote target, not whether it can wander a codebase and invent a new bug class. That is why a 48.5% score can sit next to 948 Linux root exploits without contradiction: the Linux run was a long, tool-using pipeline with scout, triage, and exploit-dev agents plus a human at the end, and ExploitGym is a graded remote-exploit exam. RemoteExploitBench is the closer cousin for known bugs in GitLab, PostgreSQL, and Jenkins, where Cyber finished 95% and Astra on Daybreak Blue would not start.

When did Mozilla ship the Firefox 159 fixes OpenAI pointed to?

Firefox 159 was announced October 27, 2026, in Mozilla Foundation Security Advisory 2026-118, with impact listed as high. Mozilla used that advisory to change process as well as code: it no longer bundles internally identified memory-safety issues into one CVE and now issues an advisory per bug. Amy Burnett of OpenAI is listed as reporter on multiple high-impact items in that file, which is how the November 5 launch can point at a browser that had already started shipping the work.

How does Daybreak Red differ from Daybreak Blue?

Blue is the verified-defender track for GPT-6 Astra, with system-level cyber guardrails relaxed so teams can validate bugs and patches without the model refusing ordinary security questions. Red is the purpose-trained Cyber track for authorized exploit development and testing, including against remote targets, and it is where GPT-6 Cyber and GPT-5.6 Cyber live. Red for Critical Infrastructure is a further group with permission to test OT and ICS systems that the default safeguards are built to refuse. Identity checks, SSO, MFA, and API IP allowlisting sit on top of all of it.

How often was the Linux pipeline simply wrong?

OpenAI said 3.1% of candidate findings assumed an impossible condition and were incorrect analysis. Most of the candidates that failed reproduction still looked like real bugs that needed extra privilege, a nasty race, a precise allocation failure, or more than the 30-hour triage budget. The 74.7% reproduction rate is the figure the company is proud of; the 3.1% is the one maintainers will want next to any automated report they have not yet replayed themselves.

OpenAI told kernel users to treat an unpatched low-privilege login as root for anyone with a serious toolkit. The model that made that sentence cheap enough to print is on a waitlist. The patches are not.

Harry is the editor of NEWS ANALYSIS. He writes across the publication's ten desks, with most of his time going to the stories where a number, a filing or a study decides the argument. His working rule is simple: read the source document before writing about it, and tell the reader plainly which parts are established and which are somebody's claim. He is responsible for the standards set out on this site's Editorial Standards and Fact Checking pages, and for correcting the record openly when the publication gets something wrong.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending