🔍 Read the full analysis: The $1,279 AI Exploit And METR's Report On A Faster Lab on ThorstenMeyerAI.com
Get the little things that make your day 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 a mid-October run of AI agents produced 948 working Linux privilege-escalation exploits, including 588 it described as previously unknown, at an average API cost of about $1,279 each. METR separately reports that Anthropic’s measured model-improvement pace accelerated after October 2025, and estimates AI contributed to much of the increase; much of the supporting evidence is not public.
OpenAI says its GPT-6 Cyber system helped produce 948 working Linux privilege-escalation exploits during a mid-October test, at an average API cost of about $1,279 per exploit. On November 5, independent evaluator METR also reported that Anthropic’s measured model-improvement pace accelerated after October 2025, with its researchers estimating that AI itself may account for much of the faster progress.
The Linux figures come from OpenAI’s account of a campaign that examined about 2.8 million lines of kernel code, or roughly 7% of the project. OpenAI said its process identified 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and generated 948 working local privilege-escalation exploits after grouping findings by root cause. It said 588 exploits involved previously unknown flaws. The cost figure is calculated from an API bill of $1,213,430 divided by the 948 exploits; it is not an independent audit of the campaign.
OpenAI named the model behind the work as GPT-6 Cyber on November 5. It describes the system as a cybersecurity-focused version of GPT-6 Astra, designed for authorized security work, including vulnerability discovery and exploit development. Access is restricted to approved organizations in the company’s Daybreak Red programme, which is intended for offensive testing such as penetration tests and red teaming. OpenAI says users face identity, device and access-control requirements, and that it can monitor, suspend or revoke access.
METR’s report addresses a separate but related development: how quickly a leading AI lab is improving its systems. The evaluator said its main measure of Anthropic’s model progress grew 1.9 times faster after October 2025 than it had before. METR estimated that AI assistance may explain about 80% of the acceleration. That estimate is an interpretation of METR’s analysis, not a directly measured share of Anthropic’s work.
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 Exploits, Faster Model Progress
The two developments point to pressure on both sides of software security. If OpenAI’s figures are accurate, AI systems can help find and turn vulnerabilities into working exploits at a cost far below the historical rewards cited for some security competitions. That does not mean every flaw is exploitable in the wild, or that the reported Linux results represent a general price for producing attacks. It does suggest that defenders may have less time to identify and patch weaknesses once tools capable of automating parts of the work become more accessible.
The risks are not limited to sophisticated attackers. OpenAI says access to GPT-6 Cyber is restricted to vetted organizations, but other research cited in the source material points to the broader challenge: Epoch AI reported in late October that a freely downloadable model, GLM 5.3, could turn published Firefox fixes into working attacks for about $1,000 per Firefox version. That result is also a reported research finding, and does not establish that all software flaws can be reproduced so cheaply. It does underscore why defenders cannot rely on access controls around one commercial model alone.
For software makers, the practical implication is an increased premium on rapid patching, secure development and coordinated disclosure. For users, keeping operating systems, browsers and applications updated reduces exposure to flaws that have already been fixed. The METR report adds a longer-term concern: if AI tools help labs improve AI systems more quickly, capabilities related to software research could advance alongside them. The report does not establish how quickly that process will continue.
Products Worth Considering
eufy Security Video Doorbell E340, Dual Cameras, 2K FHD, 8GB Local Storage
Samsung 65-Inch The Frame Pro Neo QLED 4K TV (2026) Bundle with Modern Brown Bezel, HDMI Cables, and Home Theater Guide
eufy Security Essential 7-Piece Alarm System
As an affiliate, we earn on qualifying purchases.
The Evidence Behind Both Reports
OpenAI’s Linux campaign was described in a November 4 post, before the company identified GPT-6 Cyber as the model used. The company’s November 5 announcement and system card provide the source for the model’s stated purpose, access restrictions and cybersecurity rating. OpenAI labels the model “Critical” under its own Preparedness Framework, the framework’s highest cybersecurity tier; that is OpenAI’s internal assessment, not an outside certification.
OpenAI’s account is not the only security-related evidence cited. Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes in October, according to the source material. Those advisories do not identify which OpenAI model found each bug. The source says no Linux maintainer had publicly commented on the reported kernel results at the time of publication. These distinctions matter: public acknowledgments and advisories can corroborate security work, but they do not independently verify every exploit count or attribution in OpenAI’s campaign.
METR’s findings came after an eleven-week examination of Anthropic-related data, according to the report summary. The evaluator said Anthropic redacted a majority of the evidence METR considered most informative. METR therefore presented its estimate of AI’s contribution with an explicit limitation: readers must rely on the evaluator’s judgment more than it would prefer. The report describes a measured acceleration, but its explanation for why that happened is less directly verifiable from public evidence.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, on the limits of the public evidence
What Public Evidence Cannot Verify
The 948-exploit total and the 588 previously unknown flaws are figures reported by OpenAI. The material available here does not include an independent audit of the underlying exploit set, API costs or novelty classifications. Mozilla’s advisories concern Firefox fixes and do not establish which model discovered each issue; they also do not independently validate the Linux results. Public comment from Linux maintainers was not available in the source material.
It is also unclear how many of the reported Linux exploits were suitable for use outside OpenAI’s test environment, how the vulnerabilities were disclosed and patched, or how quickly fixes reached affected distributions. A working exploit in a controlled test is not automatically a successful attack against a real-world system. The source material does not provide a complete timeline for disclosure or remediation.
METR’s 1.9-times comparison applies to its main measure of Anthropic’s model progress before and after October 2025; it should not be read as a claim that every capability improved at that rate. The estimate that AI accounts for about 80% of the acceleration is METR’s interpretation, and much of the evidence behind its analysis remains withheld. How much of the result can be independently reproduced is not clear.
Patching and Independent Checks
The immediate next questions are whether independent researchers or Linux maintainers can verify the reported findings, and what the disclosure and patch status is for the affected kernel vulnerabilities. The source material does not give a definitive public schedule for those checks or for fixes across Linux distributions. OpenAI’s Daybreak Red access is expected, according to reporting cited in the source, to reach eligible Red customers over the coming weeks; the exact rollout timetable may change.
METR’s report leaves further work for the evaluator and the public: clarify what can be disclosed, test whether its measured acceleration continues, and distinguish AI assistance from other changes in Anthropic’s research process. Until more evidence is public, the 80% figure should be treated as an estimate rather than a settled causal finding.
For organizations, the practical response does not depend on resolving every uncertainty: prioritize security updates, review exposure to affected software and maintain clear vulnerability-reporting processes. Users should install operating-system and browser updates promptly. The reports point to a faster-moving security environment, but they do not establish that a particular reader or device has been compromised.
Key Questions
What is the $1,279 figure?
It is an approximate average calculated from OpenAI’s reported $1,213,430 API bill divided by 948 working Linux privilege-escalation exploits. It is based on OpenAI’s figures, not an independent cost audit.
Were all 948 Linux flaws previously unknown?
No. OpenAI said 588 of the 948 exploits involved previously unknown flaws. The remaining count should not be described as newly discovered on the basis of the supplied figures.
Who can use GPT-6 Cyber?
OpenAI says access is limited to organizations it approves for the Daybreak Red programme, which covers offensive security testing. The company describes identity, device and access-control checks for participants.
What did METR conclude about Anthropic?
METR reported that its main measure of Anthropic’s model progress grew 1.9 times faster after October 2025 than before. It estimated that AI may account for about 80% of the acceleration, while warning that much of the supporting evidence is not public.
What should users do now?
Install available operating-system, browser and application updates promptly. The reports do not show that a particular user’s device has been attacked, but timely updates can reduce exposure to vulnerabilities once fixes are available.
Source: ThorstenMeyerAI.com
