🔍 Read the full analysis: AI's Speed And Security Signals: GPT-6 Cyber Meets The METR Report 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 its restricted GPT-6 Cyber model helped produce 948 working Linux privilege-escalation exploits in an October test, including 588 it described as previously unknown. Separately, METR estimated that Anthropic’s measured AI progress accelerated after October 2025, with AI likely contributing much of the increase; METR said key evidence was redacted.
OpenAI named GPT-6 Cyber on November 5, the model it says powered an October test that generated 948 working Linux privilege-escalation exploits, while independent evaluator METR published an assessment that Anthropic’s measured model progress has accelerated. The announcements point to two related developments: AI systems are being used to find software vulnerabilities, and AI may be helping researchers build more capable models faster.
OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra for authorized security work, including finding vulnerabilities, developing exploits and testing web applications. Access is restricted to organizations approved for its Daybreak Red programme. OpenAI’s system card says customers must meet security requirements, while the company monitors use and can revoke access. The model’s “Critical” cybersecurity rating is OpenAI’s own assessment under its Preparedness Framework, not an independent rating.
OpenAI’s account of its Linux campaign says the work scanned about 2.8 million lines of kernel code, identified 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. The resulting 948 working local privilege-escalation exploits were grouped by root cause from 1,169 exploits, according to the account. OpenAI said 588 findings were previously unknown. These figures come from the company; the supplied material does not include a public, independent audit of the full run.
OpenAI also reported finding 81 previously unknown high-severity Firefox flaws. Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes in October, but do not specify which model identified each issue. Separately, METR said its main measure of Anthropic’s model progress grew 1.9 times faster after October 2025 than before. METR estimated that AI assistance to Anthropic’s researchers may account for about 80% of the acceleration, while warning that much of its supporting evidence is not public.
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
Faster Discovery, Shorter Patch Windows
The reported Linux results suggest that AI can help scale vulnerability discovery and exploit development, but the numbers do not by themselves establish how often such systems will find usable flaws outside a controlled test. If similar capabilities become widely accessible, both defenders and attackers could identify weaknesses more quickly. OpenAI’s restricted release is one response to that dual-use risk; it also means access is currently limited to approved organizations rather than the general public.
The practical pressure falls on the time between a flaw being discovered and a fix being installed. Security teams may face more findings to verify and patch, while software vendors need reliable disclosure and remediation processes. For individual users, the immediate implication is straightforward: install security updates promptly. The source material does not show that the reported Linux findings have been exploited in real-world attacks.
METR’s assessment adds a separate concern about the pace of development. If AI tools are materially helping labs improve their own models, future capabilities could arrive faster than earlier forecasts suggested. That conclusion remains an estimate based on METR’s analysis, not a directly measured share of Anthropic’s work.
Popular right now
Ninja NC501 CREAMi XL Deluxe 11-in-1 Ice Cream and Frozen Treat Maker
Shark AV2501AE AI Robot Vacuum with XL HEPA Self-Empty Base
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
As an affiliate, we earn on qualifying purchases.
The Tests Behind the Announcements
OpenAI described its Linux campaign in a November 4 post, then named GPT-6 Cyber as the model behind it on November 5. The company says the test focused on Linux kernel code and produced working local privilege-escalation exploits in test machines. Those results should not be confused with confirmed compromises of deployed Linux systems. OpenAI’s release is its fourth cyber-focused model this year, according to the source material.
Earlier figures cited in the source offer a company-reported comparison: an OpenAI run in July found 166 Linux privilege-escalation bugs at a stated API cost of $513,690, while the October run produced 948 exploits for about $1.21 million. That works out to roughly $3,100 and $1,279 per exploit, respectively, using the reported totals. These are not independent cost benchmarks, and the campaigns’ methods and scope may differ.
METR’s report is distinct from OpenAI’s product announcement. After eleven weeks working with Anthropic data, the evaluator compared the pace of change in its main model-progress measure before and after October 2025. METR said Anthropic redacted a majority of the evidence it found most informative, limiting what outside readers can verify directly.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, on the limits of its public evidence
What the Public Cannot Verify
The Linux exploit counts and the claim that 588 findings were previously unknown are OpenAI’s figures. The source material does not provide public confirmation from Linux kernel maintainers, nor does it establish which reported findings have been patched or whether any were exploited outside testing. Mozilla’s advisories confirm credit for fixes but do not map specific bugs to GPT-6 Cyber.
METR’s estimate that AI accounts for about 80% of Anthropic’s acceleration is also uncertain. The evaluator disclosed that Anthropic redacted a majority of the evidence it considered most informative, making independent scrutiny difficult. It is not clear from the public information how much the estimate would change if the withheld material were available.
It is also unclear how quickly GPT-6 Cyber will reach all approved Daybreak Red customers, what additional validation OpenAI or software maintainers will publish, and whether the reported cost-per-exploit comparison reflects equivalent testing conditions. No evidence in the supplied material confirms real-world attacks using these findings.
Disclosure, Patching and Access
OpenAI says its Daybreak Red access is being provided to approved organizations, with broader availability to eligible Red customers expected over the coming weeks, according to reports cited in the source material. The next useful evidence would include more detail about the validation and disclosure of Linux findings, and which vulnerabilities have been fixed. Public Linux maintainer responses were not available in the material used here.
Mozilla has already published advisories crediting OpenAI for October fixes, but further information would be needed to link those fixes to particular models or tests. Security teams should follow vendor advisories and apply patches as they become available. METR’s assessment may also be updated if more evidence can be reviewed or shared; no specific follow-up date is given in the source material.
Key Questions
What is GPT-6 Cyber?
OpenAI describes it as a cyber-specialized version of GPT-6 Astra for authorized security work, such as vulnerability research and testing. Access is restricted to organizations approved for Daybreak Red.
Did GPT-6 Cyber find 948 Linux flaws?
OpenAI says an October campaign powered by the model produced 948 working Linux local privilege-escalation exploits, including 588 it called previously unknown. The figures are company-reported, and the supplied material does not include an independent audit of the full results.
Does METR’s report prove Anthropic is using AI to build AI?
No. METR estimated that AI may account for about 80% of the acceleration in its measure of Anthropic’s model progress. METR also said much of the evidence it found most informative was redacted, limiting outside verification.
Have the reported Linux exploits been used in real attacks?
The source material does not confirm real-world exploitation. OpenAI described findings and exploits produced in a test campaign; it does not establish that attackers used them against deployed systems.
What should computer users do now?
Install security updates promptly and follow advisories from operating-system and software vendors. The reports do not identify a specific user action beyond keeping systems patched.
Source: ThorstenMeyerAI.com
