🔍 Read the full analysis: A $1,279 Exploit, Faster Labs, And The Signals In GPT-6 Cyber on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI says its restricted GPT-6 Cyber model powered an October Linux-kernel exercise that produced 948 working privilege-escalation exploits, including 588 it described as previously unknown, at a reported cost of about $1,279 per exploit. Separately, METR estimates Anthropic’s model capabilities have been improving about 1.9 times faster since October 2025, while warning that much of the underlying evidence is not public.
OpenAI named GPT-6 Cyber on November 5 as the model behind an October exercise that, according to the company, generated 948 working Linux privilege-escalation exploits, including 588 flaws it said were previously unknown. The release came as independent evaluator METR reported faster model progress at Anthropic, linking much of the acceleration to AI assistance while noting that substantial supporting evidence remains confidential.
OpenAI’s November 4 account of the Linux exercise says its system scanned about 2.8 million lines of kernel code, identified and analyzed 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines, and produced 948 working local exploits. The company reported an API bill of $1,213,430, or about $1,279 per exploit. These are figures from OpenAI’s account, not an independently published audit of the full exercise.
OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra for authorized security work, including finding vulnerabilities, building exploits and testing web applications. Access is restricted to organisations approved for the company’s Daybreak Red programme, intended for offensive security testing. OpenAI says applicants must meet controls such as managed devices, multifactor authentication and least-privilege access; it can monitor, suspend or revoke access. The company lists prices of $20 per million input tokens and $100 per million output tokens for approved organisations.
Separately, METR says its analysis of Anthropic’s internal data found that its main measure of model capability improved at about 1.9 times its earlier pace after October 2025. METR estimates that AI assistance may account for about 80% of the acceleration. The evaluator also says Anthropic redacted a majority of the evidence it found most informative, limiting what outsiders can verify. The OpenAI security results and METR’s Anthropic analysis are related signals of rapid AI development, but they are separate assessments and should not be treated as a single experiment.
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 Exploit Discovery Pressures Defenders
If OpenAI’s cost and output figures hold up, they suggest that AI-assisted exploit discovery could become less expensive and faster to repeat. The reported $1,279 per working Linux exploit is a company-reported average for this exercise, not a universal price or an independently verified measure of the cost of finding any vulnerability. Even with that qualification, the scale raises practical questions for organisations responsible for maintaining widely used software.
The immediate security concern is the time between a flaw being found and a fix being installed. More effective discovery can help maintainers identify and repair defects, but working exploits can also increase the risk if vulnerabilities or attack methods spread before systems are patched. The release does not establish that the reported Linux exploits were used against real systems. OpenAI says the work was conducted for security research and that results were shared with maintainers.
For technology companies and public-sector operators, the practical implication is less about predicting a particular attack than keeping software inventories current, responding promptly to vendor advisories and applying tested security updates without unnecessary delay. For individual users, installing updates remains a straightforward defensive step. The METR report adds a broader concern: if AI tools are helping labs improve models more rapidly, the pace of change may be difficult for outside evaluators and policymakers to track using public evidence alone.
Products Worth Considering
LG UltraGear 32GS60QC-B 32-inch Curved QHD Gaming Monitor (180Hz, 1ms, 1000R, AMD FreeSync, HDR10)
Bose QuietComfort Ultra Earbuds (2nd Gen) – Wireless Noise‑Cancelling Earbuds with Mic, Immersive Audio, USB‑C Charging, Up to 6 Hours Batte
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What the Two Reports Measure
The Linux figures extend a prior OpenAI experiment described in the source material. In July, the company reported finding 166 Linux privilege-escalation bugs at a cost of $513,690, or about $3,100 per bug. OpenAI’s October account reports a larger count and lower average cost, but the runs’ totals alone do not establish a like-for-like productivity comparison: methods, scope and the definition of a counted exploit matter.
OpenAI also reported that GPT-6 Cyber found 81 previously unknown high-severity Firefox flaws. Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October, but the advisories did not identify which model found each bug. The source says no Linux maintainer had publicly commented on the exercise as of November 6. Those details mean some claims can be compared with public records, while the model-specific attribution and the full set of Linux findings remain harder to check.
METR’s work is a separate evaluation of Anthropic’s rate of progress, based on data the company provided to the evaluator. Its reported 1.9-times figure concerns the pace of improvement on METR’s main measure, not a general claim that every capability is advancing at that rate. The report’s estimate that AI accounts for about 80% of the acceleration is an interpretation based on the analysis, rather than a directly observed public benchmark result.
“the public has to rely on METR’s judgment much more than would be ideal”
— METR, on the limits of its public evidence
Independent Checks Remain Limited
The headline exploit counts, cost and novelty figures come from OpenAI. The source material does not provide a complete independent audit of the Linux exercise, a public list of all findings, or enough detail to assess whether the reported cost includes every relevant expense. It is also not clear how the company’s counting and grouping methods compare with those used in its earlier run.
Mozilla’s advisories support the claim that OpenAI was credited with numerous Firefox security fixes, but they do not connect each issue to GPT-6 Cyber. The supplied material says no Linux maintainer had commented publicly by November 6, leaving the status of maintainer review and disclosure unclear. OpenAI’s “Critical” cybersecurity rating is its own classification under its Preparedness Framework, not an external certification.
METR’s acceleration finding also carries a public-evidence limitation: the evaluator says Anthropic withheld much of what it considered most informative. Its 80% estimate is an attributed assessment, not a settled measurement of AI’s contribution. The source material does not establish whether the acceleration will continue, how it compares across other labs, or what additional evidence may become available.
Disclosure, Patching and Further Evidence
OpenAI says its Linux findings were shared with maintainers, but the source material gives no confirmed timetable for public disclosure of those flaws or for maintainers’ responses. Further detail from the company or the Linux community could help establish how many findings were accepted, fixed, or already known through other channels. OpenAI’s Daybreak Red access is also expected, according to reports cited in the source, to reach more approved customers over the coming weeks; the precise rollout schedule is not confirmed here.
Security teams should monitor official advisories from software vendors and maintainers rather than rely on unverified exploit counts. Organisations can review patch processes, confirm which systems are exposed to affected software, and prioritize updates when maintainers publish fixes. The reported exercise does not show that the vulnerabilities are being exploited in the wild, so that distinction should be preserved as new information emerges.
METR’s findings may be clarified by later reports or by the release of more supporting evidence. Until then, readers should treat its acceleration and AI-contribution figures as the evaluator’s estimates, with stated limits on external verification. The next meaningful developments are public confirmation of Linux findings and fixes, additional details about the GPT-6 Cyber evaluation, and new evidence that tests whether the reported pace of model improvement persists.
Key Questions
What did OpenAI say GPT-6 Cyber did?
OpenAI said the model powered an October Linux-kernel exercise that produced 948 working local privilege-escalation exploits, of which 588 were described as previously unknown. The figures have not been independently audited in the material provided.
How much did the Linux exercise cost per exploit?
OpenAI reported an API bill of $1,213,430. Dividing that reported bill by 948 exploits gives an average of about $1,279 per exploit. That is a company-reported average for this exercise, not a general market price.
Can anyone use GPT-6 Cyber?
No. OpenAI says access is limited to organisations it approves for its Daybreak Red programme, which is aimed at authorised offensive security testing. The company lists identity, device and access-control requirements and says it can suspend or revoke access.
What did METR find about Anthropic?
METR reported that Anthropic’s performance on its main measure improved at about 1.9 times its earlier pace after October 2025. METR estimated AI assistance may explain about 80% of the acceleration, while warning that much of the evidence it considered most informative was withheld.
Does the report show that the Linux flaws were used in attacks?
No such conclusion is established by the supplied material. It describes exploits generated during a security exercise, but does not report confirmed real-world attacks using them. The status of public disclosure and maintainer review remains unclear.
Source: ThorstenMeyerAI.com
