🔍 Read the full analysis: The $1,279 Exploit In Context: GPT-6 Cyber And The METR Report 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 its restricted GPT-6 Cyber model powered a Linux security campaign that produced 948 working privilege-escalation exploits, including 588 it said were previously unknown. Separately, METR reported that Anthropic’s measured AI research progress accelerated after October 2025, while warning that much of its supporting evidence is not public. The figures point to growing AI capabilities, but key claims remain difficult for outsiders to verify.
OpenAI said GPT-6 Cyber, a restricted model for authorized security work, powered an October campaign that produced 948 working Linux privilege-escalation exploits, including 588 flaws the company said were previously unknown. On November 5, independent evaluator METR also reported that Anthropic’s measured AI research progress has accelerated, though much of the evidence it reviewed remains confidential. Together, the announcements point to advances in both AI-assisted security research and the use of AI in developing AI; important parts of both accounts cannot yet be independently checked from public information.
OpenAI’s November 4 account of the Linux work said its pipeline examined about 2.8 million lines of kernel code, identified 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. The company said the process yielded 948 working local privilege-escalation exploits after grouping results by root cause. OpenAI later identified GPT-6 Cyber as the model behind the effort. The reported API bill was $1,213,430, or about $1,279 per exploit, based on those company-reported totals.
OpenAI’s figures are not the same as a public tally independently verified by Linux maintainers. The source material says no Linux maintainer had publicly commented on the campaign at the time of publication. 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 found each bug. That public record supports some security work by OpenAI, but does not independently establish every figure or attribution in the company’s broader claims.
METR’s report concerns a separate development. After eleven weeks of access to Anthropic data, the evaluator estimated that the pace of improvement on its main measure was 1.9 times faster after October 2025 than before. METR attributed about 80% of the acceleration to AI helping Anthropic researchers, but presented that as an estimate, not a directly observed or fully public accounting. The report says Anthropic redacted a majority of the evidence METR found most informative, limiting how much outside readers can assess for themselves.
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 Linux results, if the reported scale is accurate, illustrate how AI tools could reduce the time and expense involved in identifying and turning software flaws into working exploits. The reported cost of about $1,279 per exploit is far below the $3,100-per-exploit average OpenAI reported for a July run that found 166 Linux privilege-escalation bugs for $513,690. These are company-reported results from two runs, not a controlled comparison establishing that costs will fall at the same rate elsewhere.
Lower costs can help defenders find and fix weaknesses, but they can also make exploit development more accessible to people with harmful aims. A separate late-October study by Epoch AI, described in the source material, said a freely downloadable model called GLM 5.3 used published Firefox fixes to produce working attacks for about $1,000 per Firefox version. That result reinforces the concern that the capability is not limited to a single company’s restricted system, although the details and conditions of that experiment matter when comparing it with OpenAI’s campaign.
For organizations, the practical pressure is on vulnerability triage, testing and patching: a larger flow of findings can overwhelm teams that cannot assess and remediate them quickly. For individual users, the immediate advice remains routine but relevant: install security updates promptly. METR’s findings add a longer-term concern. If AI contributes to research that improves future AI systems, capability gains could feed into a faster development cycle, though the report’s estimate does not settle how broad or durable that effect is.
Products Worth Considering
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Black
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
EcoFlow DELTA 3 Portable Power Station (1024Wh, 1800W)
As an affiliate, we earn on qualifying purchases.
Restricted Model, Separate Evaluator
OpenAI describes GPT-6 Cyber as a cyber-specialized variant of GPT-6 Astra intended for authorized security work. It is available only to organizations approved for the company’s Daybreak Red program, which covers offensive testing such as penetration tests and red teaming. OpenAI’s listed safeguards include organizational security requirements, identity checks for individual researchers, hardware security keys, default limits on API access to approved IP addresses, usage monitoring and the ability to suspend access. The company’s developer pricing page lists $20 per million input tokens and $100 per million output tokens for approved organizations.
OpenAI rates the model “Critical” for cybersecurity, the top level in its own Preparedness Framework. That is a company assessment, not an outside certification. The source material describes GPT-6 Cyber as OpenAI’s fourth cyber-focused model released in 2026, following releases in April, June and August. The reported July Linux run provides a previous cost and output point, but does not by itself establish how much of the change came from the model, the process or other factors.
METR is an independent evaluator, but its Anthropic report was based on access to company data that is not fully available to the public. Its estimate of faster progress and the share it attributes to AI assistance should therefore be read as METR’s analysis of that evidence, rather than as a result readers can reproduce from a complete public dataset.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, on the limits of public verification
What Public Evidence Cannot Settle
The full technical details needed to independently reproduce OpenAI’s Linux results are not included in the source material. It remains unclear which of the reported vulnerabilities have been disclosed to maintainers, how many have been patched, and whether maintainers have confirmed the total. The distinction between candidate findings, reproduced vulnerabilities and working exploits also matters: the figures describe successive stages of a company-reported process, not interchangeable counts.
For Firefox, Mozilla advisories credit OpenAI with 50 high-impact fixes in October, but do not identify the model responsible for each. That leaves the claimed connection between the 81 newly found high-severity flaws and GPT-6 Cyber incompletely visible in public records. It is also not clear from the material how the October Linux cost and output figures were independently audited or how they compare under identical conditions with the July run.
METR’s report is limited by its own account of extensive redactions. The public cannot inspect most of the evidence it judged most informative, and the estimate that AI accounts for about 80% of Anthropic’s acceleration depends on METR’s interpretation of that material. The figures do not establish that the same pace will continue, or that the estimate applies to other AI labs.
Disclosure and Replication Will Matter
The next useful evidence would be public vulnerability disclosures, maintainer responses and patch records that clarify how many Linux findings were valid, new and resolved. Those records could also help distinguish the security value of AI-assisted discovery from the risks of publishing or circulating exploit details. The source material does not provide a date for a Linux maintainer response or a complete disclosure schedule.
OpenAI said access to GPT-6 Cyber is limited to approved Daybreak Red organizations, with broader access to eligible Red customers expected over the coming weeks, according to industry reports cited in the source. Whether the controls prevent out-of-scope use, and how the model performs outside OpenAI’s reported tests, remain matters for future evidence.
For METR’s findings, greater access to methods, redacted evidence or follow-up measurements would let researchers assess the acceleration estimate more directly. Until then, the report is a signal about possible AI-driven acceleration inside one lab, not a definitive measure of the field as a whole. In the nearer term, organizations should watch for verified disclosures and apply available security updates without waiting for every broader claim about AI capability to be settled.
Key Questions
What did OpenAI say GPT-6 Cyber did?
OpenAI said the model powered an October Linux campaign that produced 948 working local privilege-escalation exploits. The company said 588 involved previously unknown flaws; those totals have not all been independently confirmed in public records described by the source.
How was the $1,279 figure calculated?
It is the reported $1,213,430 API bill divided by 948 exploits, yielding about $1,279 per exploit. Both the bill and the exploit total come from OpenAI’s account of the campaign.
Can anyone use GPT-6 Cyber?
No. OpenAI says access is limited to organizations it approves for its Daybreak Red program, which is intended for authorized offensive security testing. Its listed requirements include identity and security checks, and OpenAI says it can monitor, suspend or revoke access.
What did METR conclude about Anthropic?
METR estimated that Anthropic’s progress on its main measure was 1.9 times faster after October 2025 than before. It estimated that AI assistance accounted for about 80% of the acceleration, while cautioning that much of the most informative evidence was redacted.
What should users do now?
Install security updates promptly, especially when software providers release fixes for vulnerabilities. The reports do not establish that a specific user’s device is affected, but faster vulnerability discovery makes timely patching a practical precaution.
Source: ThorstenMeyerAI.com
