🔍 Read the full analysis: METR's Faster Lab, GPT-6 Cyber, And The Cost Of An Exploit 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 GPT-6 Cyber powered a mid-October run that produced 948 working Linux privilege-escalation exploits, including 588 it says were previously unknown, at a reported API cost of $1.21 million. Separately, METR reports that Anthropic’s progress on its main measure accelerated after October 2025, while cautioning that much of the supporting evidence is confidential. Both developments point to faster AI-assisted security research, but important claims cannot yet be independently checked from public material.
OpenAI named GPT-6 Cyber on November 5 as the model behind a mid-October effort that, according to the company, generated 948 working Linux privilege-escalation exploits at an API cost of $1,213,430. On the same day, independent evaluator METR reported that Anthropic’s performance on its main measure of model progress improved about 1.9 times faster after October 2025 than before, while warning that much of the evidence is not public. Together, the reports describe faster AI-assisted vulnerability research and AI development, but the central figures are not all independently verifiable.
OpenAI’s November 4 account, later linked to GPT-6 Cyber, says its pipeline examined about 2.8 million lines of Linux kernel code, assessed 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. It reports 948 working local privilege-escalation exploits after grouping findings by root cause, with 588 flaws previously unknown. The reported API bill works out to roughly $1,279 per exploit. These are OpenAI’s figures; public Linux maintainer confirmation was not included in the source material.
OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra, intended for authorized security work such as vulnerability discovery, exploit development and web application testing. Access is restricted to organisations approved for the Daybreak Red programme, which is for offensive security testing. OpenAI’s system card lists identity checks and technical safeguards, including managed devices and restricted API access. The company rates the model “Critical” for cybersecurity, the highest tier in its own framework; that rating is not an external assessment.
METR’s report concerns a different development: the pace of Anthropic’s model improvements. It says the lab’s main progress measure grew about 1.9 times faster after October 2025 than in the period before. METR estimates that AI assistance may account for roughly 80% of the acceleration, but presents that as an estimate, not a directly established share. It says Anthropic redacted a majority of the evidence it found most informative, limiting public scrutiny of the conclusion.
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 Shorten the Patch Window
If OpenAI’s reported results hold up, the cost of generating working exploits has fallen sharply in its own successive tests. The source material compares an earlier July run, which reportedly found 166 Linux privilege-escalation bugs for $513,690, with the October run’s 948 exploits for $1,213,430. The calculations are based on company-reported counts and costs, not an independent replication. They nevertheless suggest that security teams may face a larger volume of potential vulnerabilities produced at lower cost.
That matters because finding a flaw is only one stage of defense. Maintainers and organisations still need to validate reports, develop fixes, distribute updates and persuade users to install them. If discovery accelerates faster than those tasks, the period in which vulnerable systems remain exposed could become harder to manage. The findings also have dual-use implications: tools that support authorized testing can be misused, which is why access controls and monitoring are part of OpenAI’s stated release approach.
METR’s result adds a separate concern for oversight. If AI tools are helping labs improve AI systems more quickly, external evaluators may have less time and less access to assess each advance. METR’s own caveat about confidential evidence makes transparency a central issue, not a settled detail. For computer users, the practical takeaway remains direct: install security updates promptly, while organisations should prioritize patching and incident-response capacity.
Products Worth Considering
DJI Mic 3 Wireless Microphone Kit (2 Transmitters, 1 Receiver, Charging Case)
Apple iPad mini (A17 Pro), 128GB, Wi-Fi 6E, Space Gray
Apple MacBook Air 15-inch with M5 chip, 16GB memory, 512GB SSD, and Liquid Retina display (Midnight)
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The Linux results and the METR report should not be treated as one experiment. OpenAI’s figures describe a specific security campaign and come primarily from the company’s account. The source material says Mozilla’s October public advisories credit OpenAI with 50 high-impact Firefox fixes, but do not identify which model found each issue. That provides some public evidence of security work, but it does not verify the Linux exploit count. The source also cites an Epoch AI study reporting that the freely downloadable model GLM 5.3 could turn published Firefox fixes into working attacks for about $1,000 per Firefox version; that is a separate result, not a validation of OpenAI’s Linux figures.
OpenAI says GPT-6 Cyber is available only to approved Daybreak Red organisations, with access through its API and Codex. Its programme has a separate Blue tier for defensive work, while this model is restricted to the Red tier. The source material reports a listed price of $20 per million input tokens and $100 per million output tokens for approved organisations. OpenAI has released other cyber-focused models earlier in 2026, making GPT-6 Cyber part of a continuing series rather than an isolated product announcement.
METR, an independent evaluator, examined Anthropic data over eleven weeks, according to the supplied report. Its finding is about a change in the rate of progress on its main measure, not a claim that every model capability or safety outcome improved at the same rate. The proposed role of AI in accelerating Anthropic’s research is METR’s estimate, and the report’s reliance on restricted evidence limits what outside readers can inspect.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, describing the limits of public scrutiny
Independent Checks Remain Limited
The source material does not provide an independent reproduction of OpenAI’s 948-exploit count, the claim that 588 were previously unknown, or the reported API cost. It also says no Linux maintainer had publicly commented at the time of publication. Mozilla’s advisories offer a separate, narrower public record of Firefox fixes credited to OpenAI, but do not connect particular bugs to GPT-6 Cyber or substantiate the Linux results.
METR’s estimate that AI accounts for roughly 80% of Anthropic’s acceleration also remains difficult to assess externally. The report says a majority of its most informative evidence was redacted; the precise underlying data and how the estimate was calculated are not fully available in the supplied material. It is not clear how much the result would change under different measures of progress or with access to the withheld evidence.
The security implications also depend on deployment details that are not established here: how often the model produces exploitable results outside the reported campaign, how well access controls prevent misuse, and how quickly maintainers can verify and patch the reported flaws. The available information does not establish that every reported exploit affected a widely deployed, unpatched system.
Access, Verification and Patching
OpenAI says GPT-6 Cyber access is limited to organisations it approves for Daybreak Red. The supplied reporting says broader availability to eligible Red customers is expected over the coming weeks, but it does not give a firm schedule. Further details from OpenAI, Linux maintainers or independent researchers could help clarify the reported findings and the process used to verify and disclose them.
Security teams should follow vendor and maintainer advisories, assess affected systems and apply available updates according to their risk policies. Users should install operating-system and application security updates promptly. The next important evidence will be public confirmation of the Linux findings, further technical documentation about the tests, and any additional detail METR or Anthropic provides about the data behind the acceleration estimate.
Key Questions
What did OpenAI report about GPT-6 Cyber?
OpenAI said the model powered a mid-October Linux campaign that produced 948 working privilege-escalation exploits, including 588 it described as previously unknown. Those figures have not been independently confirmed in the supplied material.
Who can use GPT-6 Cyber?
OpenAI says access is restricted to organisations it approves for the Daybreak Red programme, which covers offensive security testing. The source material says the model is offered through the API and Codex.
What did METR find about Anthropic?
METR reported that Anthropic’s main measure of model progress grew about 1.9 times faster after October 2025 than before. Its estimate that AI assistance accounts for roughly 80% of the acceleration is not a directly verified share, and much of the underlying evidence is confidential.
Are the Linux exploit numbers independently verified?
Not in the supplied source material. The totals and costs are attributed to OpenAI, and it says no Linux maintainer had publicly commented at the time. Public Mozilla advisories concern Firefox fixes and do not verify the Linux campaign.
What should organisations and users do now?
Organisations should monitor maintainer and vendor advisories, assess exposure and apply relevant patches. Users should install security updates promptly; the reports do not establish that every reported exploit affects a currently exposed system.
Source: ThorstenMeyerAI.com
