🔍 Read the full analysis: What GPT-6 Cyber And The METR Report Show About A $1,279 Exploit on ThorstenMeyerAI.com
Get movie nights 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 powered by GPT-6 Cyber produced 948 working Linux privilege-escalation exploits, including 588 it had not known about, at an API cost of about $1,279 per exploit. Separately, METR reports that Anthropic’s measured model-improvement pace accelerated, while warning that much of the evidence is confidential. The figures come mainly from company reporting, and important details remain independently unverified.
OpenAI named GPT-6 Cyber on November 5 as the model behind a mid-October Linux security run that it says generated 948 working privilege-escalation exploits, including 588 for previously unknown flaws. OpenAI put the API cost at $1,213,430, or about $1,279 per exploit; on the same day, evaluator METR reported a faster pace of model improvement at Anthropic, though much of its supporting evidence is not public.
OpenAI’s November 4 account describes a process that examined about 2.8 million lines of Linux kernel code, roughly 7% of the project. The company says its agents identified and analysed 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and produced 948 working local privilege-escalation exploits. OpenAI reports that 588 exploits concerned flaws it had not previously known about. These figures are from OpenAI’s account; the material supplied does not include an independent audit of the run or its cost.
On November 5, OpenAI identified the model as GPT-6 Cyber, a cybersecurity-focused version of GPT-6 Astra intended for authorised security work. Its system card lists vulnerability discovery, exploit development and web application testing among its uses. Access is restricted to organisations approved for the offensive-testing tier of OpenAI’s Daybreak programme, with security controls including identity checks, managed devices and limited API access. OpenAI calls the model “Critical” under its own cybersecurity preparedness framework, the framework’s highest tier; that rating is the company’s assessment, not an independent certification.
METR’s separate report says its measure of Anthropic’s model-improvement pace grew 1.9 times faster after October 2025 than before. METR estimates that AI assistance to Anthropic researchers may account for about 80% of the acceleration. It also says Anthropic withheld much of the underlying evidence, limiting what outside readers can inspect. The two developments concern different labs and different evidence: OpenAI’s reported cybersecurity results do not establish METR’s explanation for Anthropic’s progress.
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
Cheap Exploits Tighten Patch Deadlines
If OpenAI’s cost and output figures hold up, the cost of producing working exploits could become a more immediate security concern for software makers and system operators. The figure of about $1,279 per exploit is an average derived from the reported API bill and exploit count. It does not include all possible research, infrastructure or validation costs, and it does not show how easily the exploits could be used outside controlled tests.
The concern is not limited to models that are restricted to approved organisations. The source material also cites Epoch AI research from late October reporting that a downloadable model, GLM 5.3, converted published Firefox fixes into working attacks for about $1,000 per Firefox version. That is a separate result, with its own methodology; it suggests that publicly available information and models may also support attack development, but does not independently verify OpenAI’s Linux figures.
For defenders, the practical pressure is time: vulnerabilities may be found and turned into attack code more quickly, leaving less room between disclosure and exploitation. The report does not establish that these Linux exploits were used against real systems. It does reinforce the value of prompt patching, secure configuration and tested incident-response plans, while leaving the scale of actual attacker use unknown.
Popular right now
Samsung 65-Inch The Frame Pro Neo QLED 4K TV (2026) Bundle with Modern Brown Bezel, HDMI Cables, and Home Theater Guide
Shark HydroVac MessMaster Wet/Dry Vacuum & Mop Combo (AW261)
Dyson Ball Animal 3 Upright Vacuum – Corded, 290 AW, De‑tangling Motorbar Cleaner Head, 3 Suction Modes
As an affiliate, we earn on qualifying purchases.
A Restricted Model, Two Separate Reports
GPT-6 Cyber is described by OpenAI as its fourth cyber-focused model released in 2026, following releases in April, June and August. Daybreak has separate access tiers: Blue for defensive work such as code review and incident response, and Red for offensive security testing. GPT-6 Cyber is limited to organisations approved for Red access and for this model. OpenAI says it can monitor use and pause or revoke access.
The Linux result follows an earlier OpenAI run reported in July: 166 Linux privilege-escalation bugs for an API bill of $513,690, or about $3,100 each. OpenAI’s October figures imply a lower average cost per exploit, but the two runs may differ in scope and method, and the supplied account does not provide enough detail to treat them as a directly controlled comparison. The source material also cites past hacking-competition rewards of $20,000 to $70,000 for an exploit, but those rewards are not equivalent to the cost of generating one with an AI system.
For Firefox, Mozilla’s public advisories reportedly credit OpenAI with 50 high-impact fixes in October. They do not identify which model found each flaw. That public record offers some evidence of security work resulting in fixes, but it does not verify the separate claim that GPT-6 Cyber found 81 previously unknown high-severity Firefox flaws. The supplied account says no Linux maintainer had publicly commented on the kernel findings at the time of writing.
““a majority of the evidence we found most informative””
— METR, on evidence withheld by Anthropic
What the Public Can Verify
The Linux counts, the number of previously unknown flaws and the API bill are OpenAI’s figures. The supplied material does not show an independent reproduction of the exploits, a public technical inventory of the findings or a detailed accounting of what the cost includes. It is also unclear how many vulnerabilities have been patched, whether maintainers have confirmed the reported findings, or whether any exploit was used outside testing.
For Firefox, Mozilla advisories reportedly credit OpenAI with 50 high-impact fixes, but do not connect individual bugs to GPT-6 Cyber. That leaves the separate OpenAI figure of 81 previously unknown high-severity Firefox flaws without a clear public, model-by-model confirmation in the material provided. METR’s conclusions about Anthropic are also constrained by confidential evidence, and its estimate that AI accounts for about 80% of the acceleration is an assessment rather than a directly observable public measurement.
The cost comparison between the July and October Linux runs also has limits: the source gives totals and averages but not enough information to establish that the runs used identical methods or covered comparable vulnerabilities. The exploit count should not be read as a count of confirmed attacks on deployed systems.
Disclosure, Patching and Independent Review
The next key information would come from Linux maintainers and affected software teams: confirmation of the reported flaws, disclosure of patches and details about which findings have been fixed. OpenAI says GPT-6 Cyber is available only to approved Daybreak Red organisations, with broader access for eligible Red customers expected over the coming weeks, according to coverage cited in the source material.
Further public reporting could clarify the methods behind the Linux run, whether its findings have been independently checked and how the reported costs were calculated. For users and organisations, the immediate practical step remains to apply security updates promptly and follow vendor guidance; the report does not establish a specific active threat requiring a different response.
METR’s assessment may also be tested against later public evidence about model capabilities and research productivity. Until more of the underlying evidence is available, its findings are best read as a documented evaluator’s judgment with acknowledged limits, not as a fully reproducible account of Anthropic’s internal development process.
Key Questions
What did OpenAI say GPT-6 Cyber produced?
OpenAI says a mid-October run generated 948 working Linux local privilege-escalation exploits, including 588 for previously unknown flaws. The figures have not been independently verified in the supplied material.
How did OpenAI arrive at $1,279 per exploit?
OpenAI reported an API bill of $1,213,430 for the run. Dividing that amount by 948 reported exploits gives an average of about $1,279 each; the figure does not necessarily include every cost of the work.
Can anyone use GPT-6 Cyber?
No. OpenAI says access is restricted to organisations it approves for the offensive-testing tier of its Daybreak programme and for this model. The company lists identity, device and access controls for participants.
Did METR independently verify Anthropic’s internal evidence?
METR evaluated evidence from Anthropic but says the company redacted much of the material it considered most informative. METR reported a faster improvement rate and estimated AI may explain about 80% of the acceleration, while acknowledging that public readers must rely heavily on its judgment.
Does the report mean these exploits are being used in attacks?
No such use is established by the material provided. The Linux exploits were described as results from a security-testing run, and the source does not report confirmed attacks using them. The patch status and any real-world use remain unclear.
Source: ThorstenMeyerAI.com
