🔍 Read the full analysis: A $1,279 Exploit And Faster AI Research: What GPT-6 Cyber Shows 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 AI-agent run produced 948 working Linux privilege-escalation exploits, including 588 flaws it had not known about, at a reported cost of about $1,279 per exploit. The figures come from OpenAI, and public details do not independently establish which model found each flaw. A separate METR report says Anthropic’s measured model-improvement pace has accelerated, while noting that much of the supporting evidence is private.
OpenAI named GPT-6 Cyber on November 5, saying the cyber-focused model powered an AI-agent run that produced 948 working Linux privilege-escalation exploits in mid-October. OpenAI says 588 involved flaws not previously known to it, and puts the API cost at about $1,279 per exploit; those figures are company-reported, not an independent audit of the run.
OpenAI’s account describes a pipeline that scanned about 2.8 million lines of Linux kernel code, or roughly 7% of the project. The agents identified and analyzed 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines, and produced 948 working local exploits, according to the company’s post. OpenAI says the exploits were grouped by root cause and that 588 represented previously unknown flaws. The result is a measure of what the company says its system accomplished in a particular test, not proof that every Linux vulnerability was found or that the same results would carry over to other software.
The reported cost was $1,213,430 in API usage, which OpenAI’s figures work out to about $1,279 per working exploit. The source material compares that with an earlier OpenAI Linux run in July: 166 reported exploits for $513,690, or about $3,100 each. These are company-reported results and costs, and the comparison is between the two runs described by OpenAI. OpenAI also reported 81 previously unknown high-severity Firefox flaws and a 48.5% result on ExploitGym; those results, too, are attributed to the company.
OpenAI describes GPT-6 Cyber as a cyber-specialized variant of GPT-6 Astra intended for authorized security work, including vulnerability discovery, exploit development and web-application testing. Access is restricted to organizations approved for the Daybreak Red programme, which covers offensive security testing. The company says participating organizations must meet security requirements such as multi-factor authentication and managed devices. OpenAI’s “Critical” cybersecurity rating is its own assessment under its Preparedness Framework, rather than an outside 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 Costs Reshape Security Work
The reported cost matters because it suggests that producing working exploits through an AI-assisted process may become less expensive. That does not establish the cost of exploiting systems in the real world, nor does it show that all flaws can be found or weaponized this way. It does point to a possible change in the economics of vulnerability research: organizations may need to respond to a larger volume of potential findings and do so sooner.
For defenders, the practical issue is the time between a vulnerability being discovered and a fix being applied. If automated tools make it easier to identify and test weaknesses, attackers may also benefit from the same general capabilities, even when a particular model is access-restricted. The source material cites separate Epoch AI work in which a downloadable model reportedly turned published Firefox fixes into working attacks for about $1,000 per Firefox version. That is a distinct study, not evidence that GPT-6 Cyber was used in those attacks. For software users, the immediate, well-established advice remains to install security updates promptly.
The METR report adds a related but separate development. It estimates that Anthropic’s measured pace of model improvement rose after October 2025 and suggests AI assistance may account for much of that acceleration. Faster model development could bring stronger capabilities more quickly, but METR says important evidence is not public. The report therefore supports a qualified claim about measured progress, not a definitive explanation of every factor behind it.
Popular right now
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
De'Longhi Rivelia Fully Automatic Espresso Machine with Milk Frother, White
Samsung 65-Inch The Frame Pro Neo QLED 4K TV (2026) Bundle with Modern Brown Bezel, HDMI Cables, and Home Theater Guide
As an affiliate, we earn on qualifying purchases.
Restricted Access, Limited Verification
GPT-6 Cyber is presented as OpenAI’s fourth cyber-focused model release of 2026, following releases in April, June and August, according to the source material. Its access model distinguishes between Daybreak Blue, for defensive activity such as code review and incident response, and Daybreak Red, for authorized offensive testing. GPT-6 Cyber is limited to approved Red organizations. OpenAI says it can monitor usage and pause or revoke access; reports cited in the source say availability for all eligible Red customers was expected to expand over the following weeks.
Some findings have an external public record, but that does not verify every number in OpenAI’s account. Mozilla’s October advisories credit OpenAI with 50 high-impact Firefox fixes, while not identifying which model found each bug. The source says Linux maintainers had not publicly commented on the reported kernel campaign at the time of publication. The gap between company reporting and public confirmation is relevant when interpreting totals such as 948 exploits and 588 previously unknown flaws.
The two developments in the headline are related by the wider role of AI in cybersecurity and AI research, but they are not one study. OpenAI reported a security-testing campaign and named its model; METR independently assessed evidence about Anthropic’s development pace. METR’s report draws on data shared by Anthropic, with much of the evidence withheld from public view. The source author also disclosed that Anthropic’s Claude helped draft the original analysis and that Anthropic and OpenAI are subjects of it.
““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 Record Cannot Confirm
The Linux totals, the count of previously unknown flaws and the campaign’s cost are reported by OpenAI. The source material does not describe an independent audit of the full Linux run, and it says no Linux maintainer had commented publicly at the time. It is also unclear from the public information which model identified each vulnerability, how the reported findings were handled by maintainers, and how many resulted in fixes or public advisories.
Mozilla’s advisories provide a checkable record of 50 high-impact Firefox fixes credited to OpenAI, but they do not establish that GPT-6 Cyber found those bugs. The relationship between that advisory count and OpenAI’s separate claim of 81 previously unknown high-severity Firefox flaws is not fully explained in the supplied material. The ExploitGym score and the comparison with the July Linux run also depend on OpenAI’s stated methodology and definitions.
METR’s estimate that Anthropic’s improvement rate rose 1.9 times after October 2025 is an evaluator’s finding based on information from Anthropic, not a fully public dataset. METR attributes about 80% of the acceleration to AI helping Anthropic’s researchers, but presents this as an estimate. The report’s redactions leave outside readers unable to independently inspect much of the evidence behind that conclusion.
Disclosure, Patching and Access
OpenAI’s next stated step is to make GPT-6 Cyber available to approved Daybreak Red organizations over the coming weeks, according to the reports cited in the source material. Public details about additional Linux findings, maintainer responses and resulting fixes would help clarify the security impact of the campaign. The company’s restricted-access terms and monitoring approach will also matter as eligible organizations begin using the model.
For organizations responsible for Linux systems and other software, the near-term response is to follow vendor and project security advisories, apply relevant patches, and review how quickly reported vulnerabilities move through internal remediation processes. The source material does not give a timeline for any Linux fixes arising from the campaign, so readers should not assume that every reported flaw has been patched or publicly disclosed.
METR’s report may prompt further scrutiny of how AI labs measure and disclose gains from AI-assisted research. Until more evidence is public, its estimates should be read as informed but qualified findings. The central development to watch is whether security researchers and software maintainers can verify findings and ship fixes as quickly as automated systems can identify and test vulnerabilities.
Key Questions
What did OpenAI say GPT-6 Cyber produced?
OpenAI says an AI-agent run in mid-October produced 948 working Linux privilege-escalation exploits, including 588 involving flaws it had not previously known about. These figures are reported by OpenAI.
How was the $1,279-per-exploit figure calculated?
OpenAI reported an API bill of $1,213,430 for the run. Dividing that reported amount by 948 reported working exploits gives roughly $1,279 per exploit. It is a cost estimate for that run, not a general price for finding or exploiting software vulnerabilities.
Can anyone use GPT-6 Cyber?
No. OpenAI says access is restricted to organizations it approves for Daybreak Red, its programme for offensive security testing. The company says participating organizations must meet security requirements and that access may be monitored or revoked.
Have the reported Linux flaws been independently confirmed?
The supplied material does not establish independent confirmation of the complete set of Linux results. It says the totals come from OpenAI and that Linux maintainers had not publicly commented at the time. Mozilla’s advisories independently provide a record of some Firefox fixes credited to OpenAI, but do not identify which model found them.
What did METR report about Anthropic?
METR said its main measure of Anthropic’s model-improvement pace rose 1.9 times faster after October 2025 than before, and estimated that AI assistance accounts for about 80% of the acceleration. METR also said much of the evidence it considered informative was redacted, so its explanation cannot be fully checked using public material.
Source: ThorstenMeyerAI.com
