🔍 Read the full analysis: GPT-6 Cyber And The METR Report: A Snapshot Of AI's Evolving Pace on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
OpenAI named GPT-6 Cyber as the model behind a reported Linux security campaign that produced 948 working privilege-escalation exploits; the figures come from OpenAI, and independent confirmation is limited. Separately, METR reported that Anthropic’s measured model-improvement pace was about 1.9 times faster after October 2025, while cautioning that much of the evidence is redacted.
OpenAI named GPT-6 Cyber on November 5 as the model behind a mid-October campaign that the company says generated 948 working Linux privilege-escalation exploits, including 588 for previously unknown flaws. The disclosure coincided with an independent METR report estimating that Anthropic’s measured model-improvement pace rose to about 1.9 times its earlier rate; METR said much of the evidence it reviewed remains confidential.
OpenAI described the Linux work as an automated process that scanned about 2.8 million lines of kernel code, or roughly 7% of the project. Its account says agents assessed 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and produced 948 working exploits. OpenAI reported that 588 exploits involved previously unknown flaws. The API bill was $1,213,430, equivalent to about $1,279 per working exploit using the company’s figures. These are company-reported results; the supplied materials do not include an independent audit of the full run.
OpenAI calls GPT-6 Cyber a cyber-specialized version of GPT-6 Astra for authorized security work, including vulnerability discovery, exploit development and web-application testing. Access is restricted to organisations approved for its Daybreak Red programme, which is intended for offensive security testing. OpenAI’s system card assigns the model its highest internal cybersecurity risk tier, “Critical”; that designation is the company’s own assessment, not an outside certification.
METR’s separate report examined Anthropic data over 11 weeks. It estimated that Anthropic’s main measure of model progress increased 1.9 times faster after October 2025 than before. METR attributed about 80% of the current acceleration to AI helping Anthropic researchers, but characterized that share as an estimate. It also said Anthropic redacted a majority of the evidence METR considered most informative, limiting what the public can independently assess.
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 Work
Taken together, the disclosures point to two related pressures: AI systems may be helping security teams find vulnerabilities at lower reported cost, while AI-assisted research may also speed up the development of more capable models. Neither source establishes that these trends will proceed at a fixed rate, but both bear on how quickly organisations may need to respond to software risks.
OpenAI’s reported cost of roughly $1,279 per Linux exploit is far below the $20,000 to $70,000 figures the source material gives for past hacking competitions. The comparison is indicative rather than like-for-like: the supplied account does not establish that competition prizes and automated exploit-generation costs measure the same work or quality. Still, if similar capabilities spread, defenders could face a shorter period between a flaw being found and being exploited.
The practical constraint may shift toward triage, patching and deployment. A vulnerability is not eliminated when it is discovered; affected software must be assessed, fixes developed and updates installed. For users and organisations, keeping systems updated remains a direct defensive measure, while security teams may need to prepare for more findings to review.
Popular right now
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
Samsung 65 The Frame QLED 4K TV (2026) QN65LS03HEFXZA Bundle
Bose QuietComfort Wireless Noise Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Earlier Runs to Restricted Access
OpenAI’s Linux campaign follows an earlier run in July that, according to the source material, found 166 privilege-escalation bugs for $513,690—about $3,100 per bug. The October figures suggest a lower reported cost per exploit, but the source does not provide enough methodological detail to determine how directly the two runs compare.
OpenAI also said its agents found 81 previously unknown high-severity Firefox flaws. Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October, but do not identify which model found each issue. The figures therefore describe different things and should not be treated as a one-to-one confirmation of OpenAI’s model-specific claim.
The broader defensive concern is not limited to restricted commercial systems. The source material cites Epoch AI research published in late October reporting that a freely downloadable model, GLM 5.3, could turn published Firefox fixes into working attacks for about $1,000 per Firefox version. That report suggests relevant capabilities may not remain confined to a single vendor’s controlled access programme.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, on its access to Anthropic evidence
What the Public Cannot Verify
The Linux exploit count, the number of previously unknown flaws and the reported API cost are OpenAI’s figures. The supplied source material does not document an independent technical audit or provide enough detail to verify each exploit’s quality, novelty and root-cause grouping. It also says Linux maintainers had not commented publicly at the time of writing.
Mozilla’s advisories confirm credits for fixes but do not establish which model identified each flaw. The 81-flaw claim and the 50 public advisory credits therefore cannot be assumed to refer to an identical set. The source material also does not establish how quickly the Linux findings were disclosed, fixed or incorporated into released updates.
METR’s 1.9-times estimate is based on an evaluation for which it says key evidence was redacted. Its estimate that AI accounts for about 80% of acceleration is an attribution by METR, not a directly observed measure of every contribution. The precise methods, underlying data and degree to which the result applies beyond Anthropic remain unclear from the supplied material.
Access, Patching and Further Evidence
OpenAI says GPT-6 Cyber is available only to approved Daybreak Red organisations and that access is subject to identity, device and usage controls. The source material, citing media reports, says availability to all eligible Red customers is expected over the coming weeks. OpenAI can pause or revoke access, but the materials do not describe how often that will happen or what public reporting will accompany it.
The next useful evidence will include technical details on the Linux findings, confirmation from maintainers or affected projects, and information about fixes and disclosure timelines. For METR’s assessment, fuller access to the underlying evidence—or later evaluations that can be independently checked—would help clarify whether the measured pace persists. Meanwhile, organisations should track security advisories and apply relevant patches promptly.
Key Questions
What is GPT-6 Cyber?
OpenAI describes it as a cyber-specialized version of GPT-6 Astra for authorized security work, including finding vulnerabilities and testing software. Access is restricted to organisations approved for Daybreak Red.
Did OpenAI independently verify the 948 Linux exploits?
The supplied material reports the number from OpenAI’s account. It does not include an independent audit of the full campaign, and Linux maintainers had not publicly commented at the time described.
What did METR conclude about Anthropic?
METR estimated that Anthropic’s main model-improvement measure grew about 1.9 times faster after October 2025 than before. It estimated AI assistance accounted for about 80% of the acceleration, while warning that much of its informative evidence was redacted.
Are the Firefox figures independently confirmed?
Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October. They do not say which model found each issue, so they do not independently verify the claim that GPT-6 Cyber found 81 previously unknown high-severity flaws.
What should computer users do now?
Install security updates promptly and follow advisories from software and device providers. The reports do not identify a new user action beyond routine patching and security maintenance.
Source: ThorstenMeyerAI.com
