🔍 Read the full analysis: GPT-6 Cyber And METR: Exploit Costs And The Pace Of AI Research on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
OpenAI says its restricted GPT-6 Cyber model powered a Linux security run that produced 948 working privilege-escalation exploits, including 588 it described as previously unknown, at about $1,279 per exploit. Separately, METR estimates Anthropic’s model-improvement pace accelerated after October 2025, with AI likely contributing substantially; much of the evidence was redacted. Both reports point to rapid AI progress, but key claims rely on company data or limited public evidence.
OpenAI on November 5 named GPT-6 Cyber as the model behind a Linux security test that it says generated 948 working privilege-escalation exploits, including 588 previously unknown flaws, for an API cost of about $1.21 million. On the same day, independent evaluator METR reported that Anthropic’s measured pace of AI-model improvement accelerated after October 2025, with AI itself likely helping drive much of that change.
OpenAI described the Linux work as a mid-October campaign. The agents scanned about 2.8 million lines of kernel code, around 7% of the project, and investigated 2,308 candidate findings. OpenAI says the pipeline reproduced 1,724 vulnerabilities in test machines and produced 948 working local privilege-escalation exploits, after grouping findings by root cause. The figures and the count of 588 previously unknown flaws come from OpenAI’s account; they are not an independent audit of the full run.
Dividing the reported API bill of $1,213,430 by 948 gives about $1,279 per working exploit. OpenAI’s account compares this with a July run that produced 166 Linux privilege-escalation bugs for $513,690, or about $3,100 each. These are costs reported for particular runs, not a complete measure of the staffing, validation or remediation costs involved in security research.
GPT-6 Cyber is described by OpenAI as a cyber-specialized version of GPT-6 Astra for authorized security work. Access is limited to organisations approved for the company’s Daybreak Red programme, aimed at offensive testing such as penetration tests and red teaming. OpenAI’s system card gives the model a “Critical” cybersecurity rating, the highest tier in its own Preparedness Framework. That rating is the company’s assessment, not an independent classification.
METR’s separate analysis examined Anthropic data over 11 weeks. It says its main measure of Anthropic’s model-improvement pace grew 1.9 times faster after October 2025 than before. METR estimates that AI assistance may account for about 80% of the acceleration. The evaluator also says Anthropic redacted a majority of the evidence it considered most informative, leaving outside readers less able to check the underlying analysis.
GPT-6 Cyber · the Linux run
948 working exploits, about $1,279 each
OpenAI’s agents worked through the Linux kernel in mid-October. Every number on this board comes from OpenAI; our own arithmetic is marked ≈.
API cost per working Linux privilege-escalation exploit in OpenAI’s October run: $1,213,430 for 948 exploits. OpenAI’s own figures.
- Candidate findings analysed2,308
- Reproduced in test machines1,724
- Working exploits948
- …previously unknown588
About 2.8 million lines of kernel code scanned (≈7%). 948 = unique flaws after grouping 1,169 exploits by root cause.
Three months, one curve
OpenAI Linux runs, cost per privilege-escalation bug
- July 2026 · GPT-5.6 Sol166 bugs for $513,690≈ $3,094
- October 2026 · GPT-6 Cyber948 exploits for $1,213,430≈ $1,279
- Past hacking-competition payoutsper Linux privilege-escalation exploit$20k–70k
What is and is not known
As of 6 Nov 2026
- Reports go to kernel maintainers with candidate patches; a human checks each one first (OpenAI).
- Technical details, including at least 10 new techniques, are withheld until fixes ship.
- No public statement from Linux kernel maintainers found yet.
- Validated flaws with suggested patches are not the same as shipped fixes.
Sources: OpenAI, “A thousand fewer paths to root”, 4 Nov 2026, as quoted by The Hacker News and Decrypt; 588 previously unknown and the model as reported by Fortune, CSO Online and The Register, 5 Nov 2026. ≈ values: own arithmetic.
Powered by Thorsten Meyer AI
Lower Exploit Costs, Faster Patching
The reports matter because they describe gains on two parts of the technology cycle: AI systems finding and developing exploits, and AI tools potentially helping researchers improve the next generation of models. If OpenAI’s Linux figures hold up, the cost of producing working exploits in a controlled research setting has fallen between its July and October runs. That does not establish how often such capabilities can be used against real systems, but it points to a shorter window for defenders to identify and fix flaws.
The security benefit is not limited to attackers. The same ability to discover weaknesses can support code review, vulnerability research and remediation when access is controlled. But cheaper discovery may also make it easier for more actors to test software at scale. The practical pressure falls on organisations to track vulnerability reports, prioritize patches and update exposed systems promptly; finding bugs cheaply does not automatically make fixing them quick.
METR’s estimate adds a broader concern for oversight: if AI is helping labs accelerate research, progress could become harder to evaluate using publication timelines alone. Its 80% figure is an estimate, not a directly measured share of all Anthropic research. The redactions further limit how confidently outsiders can judge the result.
Products Worth Considering
Samsung Galaxy Tab A11+ (6 GB RAM, 128 GB storage, expandable) – Gray
Samsung Galaxy Tab A7 Lite 8.7 32GB Wi‑Fi Android Tablet – Long‑Lasting Battery, Compact Slim Design, Sturdy Metal Frame (US Version)
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
OpenAI’s model announcement followed a security run it had described publicly on November 4, under the title “A thousand fewer paths to root.” The company said the work focused on Linux kernel code and reported its results through its own testing pipeline. The account says the findings were shared with relevant maintainers, but the supplied material does not include a public response from Linux maintainers.
Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October, according to the source material. Those advisories do not identify which model found each issue, so they cannot establish that GPT-6 Cyber was responsible for all or any particular set of the fixes. OpenAI also reported 81 previously unknown high-severity Firefox flaws, but the relationship between that figure and Mozilla’s public tally is not specified.
The source material also cites an earlier Epoch AI report on a freely downloadable model, GLM 5.3, which it says converted published Firefox fixes into working attacks for about $1,000 per Firefox version. That is a separate finding and should not be treated as verification of OpenAI’s Linux results. Across these reports, benchmark scores and company-run tests offer evidence of capability, but public access to methods, samples and independent reproduction remains uneven.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, describing the limits of its review
What the Public Cannot Verify
The Linux exploit totals, previously unknown flaw count and API bill are OpenAI-reported figures. The source material does not provide an independent reproduction of all 948 exploits, a public inventory of the findings or a detailed account of how each was validated and disclosed. It is also unclear how many flaws have been patched, how many remain under coordinated disclosure, or whether maintainers have publicly assessed the results.
The Firefox figures are not directly reconciled: Mozilla’s advisories credit OpenAI with 50 high-impact fixes, while OpenAI is reported as claiming 81 previously unknown high-severity flaws. The available material does not explain the difference or attribute specific findings to GPT-6 Cyber. OpenAI’s “Critical” rating is also an internal framework judgment rather than an outside evaluation.
METR’s analysis is independent, but its conclusion depends on evidence that Anthropic did not make fully available. The report’s 80% estimate of AI’s contribution is an inference, and the source material does not provide enough public detail to assess the estimate’s uncertainty or reproduce it. Neither report establishes how quickly these capabilities will spread beyond restricted or research settings.
Disclosure, Patches and Further Review
OpenAI says it shared Linux findings with relevant maintainers, but the supplied material gives no public timetable for disclosures or patches. Public advisories and maintainer statements could clarify which flaws were confirmed, how many have been fixed and whether the reported exploits correspond to distinct vulnerabilities.
OpenAI’s Daybreak Red access is expected, according to coverage cited in the source material, to expand to approved programme customers over the coming weeks. The company says access is monitored and may be paused or revoked. Further details about deployment, safeguards and independently tested performance would help clarify how the model operates outside OpenAI’s reported run.
METR’s report leaves open whether Anthropic or the evaluator will publish additional evidence that can be checked without exposing sensitive information. For software users and administrators, the immediate response remains practical: install security updates promptly, maintain supported systems and follow advisories from software vendors. The next meaningful measure will be how quickly reported vulnerabilities are verified and repaired—not just how quickly AI can find them.
Key Questions
What did OpenAI say GPT-6 Cyber did?
OpenAI says the model powered a mid-October Linux security run that produced 948 working local privilege-escalation exploits, including 588 flaws it described as previously unknown. Those numbers come from OpenAI and have not been independently verified in the supplied material.
How much did the reported Linux run cost?
OpenAI reported an API bill of $1,213,430, equivalent to about $1,279 per exploit when divided by its reported total of 948. That figure covers the reported API cost, not all research, validation or patching expenses.
Can the public use GPT-6 Cyber?
No general public release is described. OpenAI says access is restricted to organisations it approves for Daybreak Red, its programme for offensive security testing.
What did METR find about Anthropic?
METR reported that its main measure of Anthropic’s model-improvement pace grew 1.9 times faster after October 2025 than before. It estimated that AI assistance may explain about 80% of the acceleration, while noting that substantial evidence was redacted.
What should organisations and users do now?
Follow security advisories, install available updates promptly and keep systems supported. The reports do not show that every system is exposed to these specific exploits, but faster vulnerability discovery can make timely patching more consequential.
Source: ThorstenMeyerAI.com
