🔍 Read the full analysis: The METR Report's Faster Lab And GPT-6 Cyber's $1,279 Exploit on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI says a mid-October Linux kernel test produced 948 working privilege-escalation exploits, including 588 it had not previously known, at an API cost of about $1,279 per exploit. The company named GPT-6 Cyber as the model behind the work on November 5. METR, in a separate report, estimated that Anthropic’s model capabilities are improving about 1.9 times faster than before, while warning that much of its supporting evidence is confidential.
OpenAI said on November 5 that its restricted cybersecurity model, GPT-6 Cyber, powered an October test that generated 948 working Linux privilege-escalation exploits, 588 of them described as previously unknown. The company put the API cost at about $1,213,430, or roughly $1,279 per exploit; that figure and the vulnerability counts come from OpenAI, not an independent audit. On the same day, evaluator METR reported that Anthropic’s measured AI progress had accelerated, offering a separate sign of how AI systems may be advancing both software testing and AI research.
OpenAI’s November 4 account of the Linux work said its system scanned about 2.8 million lines of kernel code, roughly 7% of the project. It identified 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and produced 948 working local privilege-escalation exploits. OpenAI also reported grouping 1,169 exploits by root cause and said 588 of the final 948 were previously unknown. The company identified GPT-6 Cyber as the model behind the run the following day.
The reported cost per exploit is lower than in an earlier OpenAI run. The source material says a July test found 166 Linux privilege-escalation bugs for $513,690, around $3,100 each; comparing those company-reported figures with October’s gives the stated decrease to about $1,279. OpenAI’s account is not an independently verified cost study, and the available material does not detail all costs beyond the API bill.
OpenAI describes GPT-6 Cyber as a cybersecurity-specialized version of GPT-6 Astra for authorized security work, including vulnerability discovery, exploit development and web-application testing. Access is limited to organisations approved for the company’s Daybreak Red programme, which is intended for offensive security testing. OpenAI says approved users face identity, device and access controls, and that it monitors use and can suspend access. Its own Preparedness Framework assigns the model a “Critical” cybersecurity rating, the framework’s highest tier; that is OpenAI’s assessment, not an external 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
Lower Exploit Costs Tighten Patch Timelines
If the reported results hold up, the Linux test points to a sharp drop in the cost of producing working exploits with AI assistance. That matters to security teams because the same capability can help authorised researchers find flaws before attackers do, while also making it easier for others to turn vulnerability knowledge into attacks. A low reported cost is not proof that exploitation is easy in real-world conditions: the figures describe a particular test environment and a company-run process.
The practical concern is the time between a flaw becoming known and systems being patched. Security teams may face more findings to review and less time to respond, while software maintainers need clear, reproducible reports and coordinated disclosure. For organisations and individual users, keeping software updated remains a direct way to reduce exposure, although the source material does not establish that the reported Linux exploits are being used in attacks.
METR’s report adds a second, related issue: AI may be helping labs accelerate their own research and development. METR estimated that Anthropic’s main progress measure grew about 1.9 times faster after October 2025 than before, and attributed about 80% of the acceleration to AI helping the company’s researchers. That estimate is METR’s interpretation of evidence it reviewed, not a directly observable measurement of every internal contribution.
Products Worth Considering
Dell XPS 13 9350 Copilot+ Laptop, 13.4-inch 120Hz Display, Intel Core Ultra 7 256V, 16GB RAM, 512GB SSD, Ubuntu Linux 24.04 LTS
Apple iPad mini (A17 Pro), 128GB, Wi-Fi 6E, Space Gray
Dell XPS 13 9345 Laptop, 13.4-Inch FHD+ 120Hz, Snapdragon X Elite, 32GB RAM, 1TB SSD
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The OpenAI and METR developments concern different organisations and rely on different kinds of evidence. OpenAI’s Linux figures describe a specific test and its stated API spending. The source material also reports that Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October, but those advisories did not identify which model found each flaw. They therefore do not independently establish that GPT-6 Cyber discovered all 50.
OpenAI released GPT-6 Cyber through a restricted programme rather than general access. The supplied account says only approved Daybreak Red organisations can use it, with access expected to reach eligible customers over the following weeks. OpenAI also lists approved-organisation API rates of $20 per million input tokens and $100 per million output tokens. These controls and prices help define the release, but do not by themselves verify the model’s reported capabilities.
METR’s assessment is an independent evaluation of Anthropic’s development pace, based on access to company data over eleven weeks. It concluded that the organisation’s measured progress had accelerated and estimated that AI assistance accounted for most of that change. METR also said Anthropic redacted a majority of the evidence it found most informative, limiting how fully outside readers can check the conclusion.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, on its access to Anthropic evidence
Independent Verification Remains Limited
The reported Linux exploit count, the number said to be previously unknown and the API bill are OpenAI’s figures. The source material does not provide a public independent reproduction of the full test, the underlying exploit set or a detailed accounting that would let readers verify the per-exploit cost. It also does not say whether all 948 exploits were disclosed to maintainers or how many have been fixed.
Mozilla’s advisories credit OpenAI with 50 high-impact Firefox fixes in October, but do not attribute individual discoveries to GPT-6 Cyber. The supplied material says no Linux kernel maintainer had commented publicly at the time of publication. It is also unclear whether any of the reported Linux findings have been exploited outside testing.
METR’s conclusion about Anthropic’s pace is qualified by limited access to the most informative underlying evidence. The 1.9-times figure describes METR’s main measure over specified periods; it should not be read as a claim that every Anthropic model capability, product or research task improved at exactly that rate. The source material does not include enough public detail to independently reproduce METR’s estimate.
Disclosure, Access and Further Evidence
OpenAI says GPT-6 Cyber is restricted to approved Daybreak Red organisations, with availability for eligible customers expected to expand over the coming weeks. The next evidence to watch is whether OpenAI or affected maintainers publish further detail on the Linux findings, including disclosure status, fixes and reproducible test results. Public documentation could clarify how the reported exploits were validated and how the stated API cost was calculated.
Readers can also look for additional Mozilla and Linux security advisories that connect specific flaws to researchers or models. Unless those records make the link explicit, the Firefox fixes should not be treated as proof that GPT-6 Cyber found each one. No timetable for public release of the Linux findings is specified in the source material.
METR’s report leaves open whether Anthropic or the evaluator will publish more evidence that can be independently examined. Until then, the pace estimate remains a reported assessment with a stated evidence limitation. For software users and administrators, the immediate response is practical: apply security updates promptly and follow advisories from the maintainers of the systems they use.
Key Questions
What did OpenAI report about the Linux test?
OpenAI said its mid-October work produced 948 working local privilege-escalation exploits, including 588 it described as previously unknown. The counts have not been independently verified in the supplied material.
How was the $1,279-per-exploit figure calculated?
It is the reported $1,213,430 API bill divided by 948 exploits, rounded to about $1,279 each. It reflects OpenAI’s stated API cost, not necessarily every cost involved in the work.
Who can use GPT-6 Cyber?
OpenAI says access is limited to organisations it approves for its Daybreak Red programme, intended for offensive security testing. The company says access is controlled and monitored.
What did METR say about Anthropic?
METR reported that Anthropic’s main measured progress grew about 1.9 times faster after October 2025 than before. It estimated that AI assistance accounted for about 80% of the acceleration, while noting that much of its most informative evidence was withheld from public view.
Does the report show that the Linux exploits are being used in attacks?
No. The supplied material describes exploits produced in a test and does not establish that they have been used in real-world attacks. It also does not specify how many reported vulnerabilities have been fixed.
Source: ThorstenMeyerAI.com
