🔍 Read the full analysis: The $1,279 Exploit And The METR Report's Twice-as-Fast Lab 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 a mid-October Linux kernel exercise produced 948 working privilege-escalation exploits, including 588 flaws it had not previously known, at an API cost of about $1,279 per exploit. Separately, METR reported that Anthropic’s model-development pace has accelerated, estimating that AI may account for much of the change. Both findings rely partly on information the public cannot independently inspect.
OpenAI said its AI agents produced 948 working Linux privilege-escalation exploits in a mid-October exercise, at an estimated API cost of about $1,279 per exploit, while independent evaluator METR reported that Anthropic’s measured model-development pace has accelerated. The paired developments matter because they point to AI both lowering the cost of finding software weaknesses and helping AI labs improve their systems faster, though key evidence is not public.
OpenAI described the Linux campaign in a November 4 post and identified the model behind it as GPT-6 Cyber the following day. The company says the agents scanned about 2.8 million lines of kernel code, or roughly 7% of the project; assessed 2,308 candidate findings; reproduced 1,724 vulnerabilities in test machines; and generated 948 working local privilege-escalation exploits. OpenAI says 588 of the 948 were previously unknown. These figures come from the company, not an independently published audit of the run.
The reported API bill was $1,213,430, which works out to roughly $1,279 per exploit using OpenAI’s count. A previous OpenAI exercise in July reportedly found 166 Linux privilege-escalation bugs for $513,690, or about $3,100 each. The comparison suggests lower reported cost per result, but it does not by itself establish that the runs were directly comparable in scope or methodology.
OpenAI describes GPT-6 Cyber as a cybersecurity-focused version of GPT-6 Astra for authorized security work, including vulnerability research and testing. Access is restricted to organisations approved for the company’s Daybreak Red programme. OpenAI’s system card rates it “Critical,” the top tier in its own preparedness framework; that is the company’s internal rating, not an independent certification.
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 Tighten Patch Windows
If the reported results hold up, a capable AI system can help produce working exploits at a per-result cost far below the amounts cited for past hacking competitions. That changes the practical scale of vulnerability research: security teams may be able to find weaknesses faster, but attackers may also gain tools for searching and exploiting them. The Linux count alone does not show that the exploits have been used outside the test setting.
The immediate pressure falls on software maintainers and organisations that rely on their code. When discovery becomes faster, the time between a vulnerability being found and a fix reaching users becomes more consequential. Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes in October, but do not identify which model found each flaw. The source material says no Linux maintainer had publicly commented on the reported campaign at publication time.
METR’s report raises a connected concern about the pace of AI development itself. It says Anthropic’s main measure of progress grew 1.9 times faster after October 2025 than before. METR estimates that AI assistance to Anthropic’s researchers may explain about 80% of the acceleration. That is an evaluator’s estimate, not a directly observable breakdown of the company’s work.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition, 32GB (2024), Metallic Black
EcoFlow Delta 3 Portable Power Station – 1024 Wh, 1800 W
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
Two Findings, Different Evidence
The two developments are related but rest on different kinds of evidence. OpenAI’s exploit counts, cost figures and model attribution are company-reported. Mozilla’s advisories offer a public check that OpenAI was credited with a set of Firefox fixes, but do not verify the Linux totals or connect particular bugs to GPT-6 Cyber. The source material also cites Epoch AI’s late-October demonstration that a 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 test and should not be treated as confirmation of OpenAI’s results.
METR’s assessment followed eleven weeks inside Anthropic’s data, according to the report described in the source material. It compares the rate of change in its principal measure before and after October 2025 and offers an estimate of how much AI contributed to the acceleration. Its conclusion is not a claim that every part of Anthropic’s research is now automated, or that the same pace applies across other labs.
OpenAI has released several cyber-focused models in 2026, according to the source material, with GPT-6 Cyber described as its fourth such model that year. The programme’s restrictions are intended to limit access to vetted organisations and authorized work. They do not, on their own, establish how well those controls will prevent misuse or how quickly the model will become available to approved users.
““a majority of the evidence we found most informative””
— METR, discussing limitations in its report
The Evidence Readers Cannot See
The public cannot independently verify the full Linux exercise from the figures provided. The available account does not include a complete list of the reported kernel findings, an outside replication of the 948 exploits, or a public response from Linux maintainers. OpenAI’s claim that 588 flaws were previously unknown also needs to be kept distinct from confirmation that each was newly discovered and handled in a particular way.
METR says Anthropic redacted a majority of the evidence it considered most informative. The report’s estimate that AI contributed about 80% of the acceleration is therefore uncertain and dependent on METR’s access and analysis. The source material does not establish how the estimate was calculated in enough public detail for readers to reproduce it.
It also remains unclear how broadly GPT-6 Cyber will be available, whether further independent security evaluations will be published, and whether any of the reported Linux exploits have been used outside controlled testing. Neither the cost-per-exploit figure nor the reported acceleration should be treated as a guarantee of future results.
More Verification and Wider Access
OpenAI says GPT-6 Cyber is available through its API and Codex to organisations approved for Daybreak Red. Reports cited in the source material say it is expected to reach all approved Red customers over the coming weeks, but the precise schedule and the number of eligible organisations are not stated. Further details from OpenAI, Linux maintainers or independent researchers could clarify the scope and handling of the kernel findings.
For organisations and individual users, the practical step supported by this reporting is to apply security updates promptly and follow advisories from the software projects they rely on. The key next test is whether maintainers confirm and address findings at a pace that matches the reported growth in automated discovery. METR’s claims about Anthropic will also be easier to judge if more of the supporting evidence becomes available or other evaluators reach comparable findings.
Key Questions
What did OpenAI say its Linux exercise found?
OpenAI reported 948 working local privilege-escalation exploits from a mid-October exercise, including 588 flaws it said were previously unknown. The figures have not been independently verified in the material available here.
What does the $1,279 figure represent?
It is an approximate per-exploit cost calculated from OpenAI’s reported $1,213,430 API bill and 948 reported working exploits. It does not include a broader accounting of research, infrastructure or security costs.
Who can use GPT-6 Cyber?
OpenAI says access is limited to organisations it approves for the Daybreak Red programme, which covers authorized offensive security testing. The company lists identity, device and access-control requirements for the programme.
What did METR conclude about Anthropic?
METR reported that Anthropic’s main measured progress grew 1.9 times faster after October 2025 than before. It estimated that AI assistance may account for about 80% of the acceleration, while acknowledging that much of the evidence it found informative was redacted.
Are the Linux exploits confirmed to have been used in attacks?
The source material describes exploits produced in a test exercise, not confirmed real-world attacks. It does not establish whether any were used outside controlled testing.
Source: ThorstenMeyerAI.com
