🔍 Read the full analysis: A $1,279 Exploit And A Faster Lab: Two AI Trends In Focus on ThorstenMeyerAI.com
Get smart everyday buys 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 an October AI-assisted Linux kernel test produced 948 working privilege-escalation exploits, including 588 flaws it had not previously known, at an average API cost of about $1,279 per exploit. Separately, METR estimates Anthropic’s model-development pace accelerated after October 2025, with AI likely driving much of the change; much of the underlying evidence is not public.
OpenAI says an AI-agent run against the Linux kernel in mid-October produced 948 working privilege-escalation exploits, including 588 flaws it had not previously known, for an API bill of $1,213,430. On November 5, the company identified the model behind the work as GPT-6 Cyber; the same day, independent evaluator METR reported that Anthropic’s model-development pace had accelerated, with AI likely contributing much of the increase.
The Linux figures come from OpenAI’s account of its own security exercise, not a public independent replication. The company said its pipeline scanned about 2.8 million lines of kernel code, roughly 7% of the project, assessed 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. It reported 948 working local privilege-escalation exploits after grouping findings by root cause; 588, OpenAI said, had not been known before.
The reported API cost works out to about $1,279 per exploit, based on the full $1,213,430 bill divided by 948. OpenAI said the run took place in mid-October and later attributed it to GPT-6 Cyber. The model is a cyber-specialized version of GPT-6 Astra, according to the company, and is restricted to organizations it approves for the Daybreak Red program, which covers authorized offensive security testing.
METR’s separate assessment concerns the rate of progress inside Anthropic, not the Linux test. After an 11-week review of Anthropic data, the evaluator said its main measure of model capability grew 1.9 times faster after October 2025 than before. METR estimated that AI itself probably accounts for about 80% of the acceleration. That estimate is an interpretation of evidence reviewed by METR, rather than a fully public dataset readers can independently inspect.
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
Cheaper Exploit Discovery, Faster Development
Together, the reports point to two ways AI may change cybersecurity: models can help find and turn software flaws into working attacks, while AI tools may also speed the research that produces more capable models. The immediate concern for system operators is the time between a vulnerability becoming discoverable and a fix being installed. If exploit development becomes less expensive, organizations may have less time to patch before others can reproduce the work.
That does not mean the reported Linux exploits were used against real systems. OpenAI described a controlled research exercise, and the supplied material does not report real-world attacks linked to the findings. Still, the cost figure is relevant to defenders because it suggests that producing working exploits may require less expenditure than in earlier efforts. A separate July OpenAI run reportedly generated 166 Linux privilege-escalation bugs for $513,690, or about $3,100 per exploit. Those are company-reported figures, not an independently standardized cost comparison.
The potential benefits and risks are closely connected. Security teams can use AI to identify vulnerabilities and test fixes, but the same capabilities could be used to develop attacks. The source also cites Epoch AI’s late-October finding that the downloadable model GLM 5.3 could turn published Firefox fixes into working attacks for about $1,000 per Firefox version. That result, too, needs to be understood as a reported research finding rather than evidence of a particular attack in the wild.
Products Worth Considering
Dell XPS 13 9350 Copilot+ Laptop, 13.4-inch 120Hz Display, Intel Core Ultra 7 256V, 16GB RAM, 512GB SSD, Ubuntu Linux 24.04 LTS
HP OfficeJet 250 Wireless Mobile All-in-One Inkjet Printer
FLEXISPOT 36-Inch Standing Desk Converter, Black
As an affiliate, we earn on qualifying purchases.
What OpenAI and METR Reported
OpenAI announced GPT-6 Cyber on November 5, describing it as a model for authorized security work, including vulnerability discovery, exploit development and web-application testing. The company’s system card assigns it a “Critical” cybersecurity rating, the highest tier in OpenAI’s own Preparedness Framework. That is a company assessment, not an outside certification. Access is limited to approved Daybreak Red organizations, with stated controls including identity checks, hardware security keys and restricted API access.
The Linux test figures were described in an OpenAI post on November 4. The source material also says Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October. Those advisories do not identify which model found each issue, so they do not independently verify that GPT-6 Cyber discovered them. As of the source report, no Linux kernel maintainer had commented publicly on OpenAI’s claims.
METR’s report offers a different kind of evidence: an external evaluator’s analysis of Anthropic’s internal data. METR said Anthropic redacted “a majority of the evidence we found most informative,” limiting what outsiders can check. The evaluator cautioned that the public must rely on its judgment more than it would ideally prefer. The two developments therefore differ both in subject and in how much supporting information is available for public review.
““a majority of the evidence we found most informative””
— METR, on access to Anthropic’s evidence
Independent Checks Remain Limited
The main Linux exploit counts and cost figures are OpenAI’s own reported results. The source material does not provide an independent reproduction of the full test, public technical details for every finding or a response from Linux kernel maintainers. It also does not establish whether all reported flaws have been fixed or when fixes became available.
Mozilla’s advisories confirm that OpenAI was credited with 50 high-impact Firefox fixes in October, but they do not say which model found the issues. That prevents those advisories from confirming GPT-6 Cyber’s specific contribution. The relationship between those Firefox fixes and the separate Linux exercise is also not established in the material.
METR’s acceleration estimate has a different limitation: much of its most informative evidence was redacted. Its 1.9-times figure and estimate that AI explains about 80% of the acceleration should be attributed to METR, not presented as independently settled facts. The source does not provide enough public detail to test those estimates directly.
Patch Responses and Further Evidence
OpenAI’s announced access plan is for GPT-6 Cyber to be available to approved Daybreak Red organizations through the API and Codex. The source material says access was expected to expand to Red customers over the following weeks, but does not establish how many organizations will receive it or what additional safeguards may be added.
For the Linux findings, the next useful developments would be public responses from kernel maintainers, technical confirmation of the vulnerabilities and information on fixes and disclosure timelines. Those details would help clarify how the reported findings affect users and whether the test led to changes in the kernel.
For Anthropic’s development pace, further public evidence or later METR assessments could show whether the reported acceleration continues and how much AI contributes. In the meantime, the practical guidance for organizations and individual users is straightforward: apply security updates promptly, restrict access to critical systems and treat the reported exploit totals as a reason to take patching seriously, not as evidence that a particular system has been attacked.
Key Questions
What did OpenAI say its Linux test produced?
OpenAI reported 948 working local privilege-escalation exploits from a mid-October test, including 588 flaws it said were previously unknown. The figures have not been independently reproduced in the source material.
How was the $1,279-per-exploit figure calculated?
It is the reported $1,213,430 API bill divided by 948 exploits, giving about $1,279 per exploit. It is an average based on OpenAI’s reported costs and output, not a general price for developing an exploit.
Who can use GPT-6 Cyber?
OpenAI says access is limited to organizations it approves for its Daybreak Red program, which supports authorized offensive security testing. The company describes identity, device and access controls for approved users.
What did METR conclude about Anthropic?
METR said its main measure of model capability grew 1.9 times faster after October 2025 than before and estimated that AI probably accounts for about 80% of the acceleration. Much of the evidence it considered most informative was not made public.
Do the reports show that these exploits were used in attacks?
No. The source material describes OpenAI’s work as a security research exercise and does not report real-world attacks tied to the Linux findings. It also does not establish whether every reported vulnerability has been fixed.
Source: ThorstenMeyerAI.com
