🔍 Read the full analysis: GPT-6 Cyber And METR: What Two AI Signals Can—and Can't—Tell Us 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 GPT-6 Cyber powered an October campaign that produced 948 working Linux privilege-escalation exploits, including 588 flaws it says were previously unknown. Separately, evaluator METR estimates Anthropic’s research pace has accelerated, with AI likely contributing substantially—but much of the evidence is confidential. The reports point to faster vulnerability discovery and AI-assisted model development, while leaving key results difficult to independently verify.
OpenAI named GPT-6 Cyber on November 5 as the model behind an October campaign that, according to the company, produced 948 working Linux privilege-escalation exploits, including 588 for previously unknown flaws. On the same day, independent evaluator METR reported that Anthropic’s measured AI research progress had accelerated, with its analysis suggesting that AI itself may account for much of the increase. The findings offer two signals of faster AI capability, but much of the underlying evidence is not public.
OpenAI said its agents scanned about 2.8 million lines of Linux kernel code, around 7% of the project, and examined 2,308 candidate findings. The company said it reproduced 1,724 vulnerabilities in test machines and produced 948 working local privilege-escalation exploits. OpenAI reported that 588 of those exploits involved previously unknown flaws. These figures come from OpenAI’s account; the source material does not include a public, independent audit of the full run.
OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra, trained for vulnerability discovery, exploit development and web application testing. Access is limited to organisations approved for the offensive-testing tier of its Daybreak Red programme. The company says those organisations must meet security requirements including multi-factor authentication, managed devices and restricted API access. OpenAI rates the model “Critical” for cybersecurity under its own Preparedness Framework, the framework’s highest tier; that rating is the company’s assessment, not an independent certification.
METR’s report examined Anthropic’s internal research data over an eleven-week period. It concluded that the organisation’s main measure of progress grew about 1.9 times faster after October 2025 than before, and estimated that AI may explain roughly 80% of the acceleration. METR also cautioned that Anthropic withheld much of the evidence it considered most informative. The estimate is an evaluator’s interpretation of restricted data, not a publicly reproducible measurement of every step in Anthropic’s research process.
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, Faster Lab Research
Taken together, the reports describe a feedback loop with consequences beyond AI labs: systems may be helping researchers build more capable systems while also making it cheaper to discover and exploit software weaknesses. The two claims are distinct, and neither proves that AI progress will continue at the same rate. But if the reported cybersecurity results are reproducible, defenders may face more findings and less time to address them.
The source material compares OpenAI’s reported October Linux run with a July run: 948 exploits for about $1.21 million, versus 166 for $513,690. Those figures imply a reported average cost of roughly $1,279 per October exploit, compared with about $3,100 in July. They are costs and counts from company-reported runs, not a market-wide measure of exploit development. The material also cites an Epoch AI experiment in which a downloadable model reportedly turned published Firefox fixes into working attacks for about $1,000 per Firefox version. That separate result suggests the capability is not necessarily limited to restricted commercial systems, though it does not establish how easily attacks could be used against unpatched systems.
For businesses, the practical pressure is on vulnerability triage, software maintenance and patch deployment. For users, the direct takeaway remains straightforward: install security updates promptly. Finding a flaw is not the same as exploiting it against a real target, and the source material does not report a wave of attacks caused by these findings. Still, lower discovery costs could make slow patching riskier.
Popular right now
Dyson V15 Detect Cordless Vacuum
Samsung QN65LS03FAFXZA 65 The Frame QLED 4K Art Mode Vision AI Smart TV (2025)
Dyson V8 Cyclone Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
What OpenAI and METR Measured
The Linux figures refer to a specific campaign, not a count of all flaws in Linux. OpenAI said its pipeline moved from scanning code to candidate findings, reproduced vulnerabilities, and then working local privilege-escalation exploits. Such an exploit could let an ordinary user on a Linux machine gain administrator privileges under the relevant conditions. The source material says the findings were shared with maintainers, but it reports no public comment from a Linux maintainer at the time of publication.
OpenAI’s Firefox work is a separate strand. Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October, while the source material says OpenAI reported 81 previously unknown high-severity Firefox flaws. The advisories do not identify which model found each issue, so the public record cannot establish that GPT-6 Cyber was responsible for every credited fix. Keeping these counts separate matters: they describe different records and do not necessarily measure the same set of vulnerabilities.
METR’s finding concerns Anthropic’s internal research productivity, not a direct benchmark of a model’s general intelligence or a forecast for the whole industry. The organisation estimated the change in its chosen progress measure and then assessed possible causes. Its estimate that AI accounts for about 80% of the acceleration is a causal interpretation, and the limited public evidence makes it harder for outsiders to check independently.
“The public has to rely on METR’s judgment much more than would be ideal.”
— METR, describing the limits of its evidence
The Evidence Still Has Gaps
The key Linux totals are based on OpenAI’s account. The source material does not provide an independent technical review of all 948 exploits, public details for each flaw, or a public response from Linux maintainers. It also does not establish whether every reported exploit remains effective outside the company’s test machines or under what exact system conditions each one works.
For Firefox, Mozilla’s advisories credit OpenAI for fixes but do not connect individual bugs to a particular model. The relationship between the 81 flaws cited by OpenAI and the 50 high-impact fixes credited in Mozilla’s public advisories is not fully explained in the source material. Nor is there enough public information to confirm the separate Epoch AI result’s broader applicability beyond its reported test conditions.
METR’s report is also constrained by access: it says much of the evidence it found informative was redacted. The report’s estimate of AI’s contribution depends on METR’s methods and judgment, while outsiders cannot inspect much of the underlying Anthropic data. It remains unclear how much of the reported acceleration will persist, whether it applies across other labs, and how quickly defensive capacity can keep pace.
Patch Releases and Further Scrutiny
OpenAI says it shared Linux findings with maintainers, but the source material does not specify a public schedule for fixes or list affected kernel versions. The next checkable developments are maintainer disclosures, security advisories and patches that clarify which issues were confirmed and when users should update. Organisations using Linux should follow their vendors’ security notices rather than assume every reported finding applies to their systems.
OpenAI has said access to GPT-6 Cyber is restricted to approved Daybreak Red organisations, with wider availability to eligible Red customers expected over the coming weeks, according to reports cited in the source material. Its deployment and monitoring practices will be relevant to judging how tightly the model is confined to authorized testing. More public technical detail or outside validation could also help establish how well the reported performance generalizes.
For METR’s assessment, further reports or disclosed evidence may clarify its measurement and the role of AI in Anthropic’s research process. Until then, the two announcements should be read as significant but qualified signals: one based mainly on a company’s cybersecurity results, the other on an evaluator’s analysis of partly confidential data. Both point toward faster capability development; neither settles how broad, durable or independently verifiable that change is.
Key Questions
What did OpenAI say GPT-6 Cyber accomplished?
OpenAI said the model powered an October Linux campaign that produced 948 working local privilege-escalation exploits, including 588 for flaws it said were previously unknown. The totals have not been independently audited in the source material.
Can anyone use GPT-6 Cyber?
No. OpenAI says access is restricted to organisations it approves for the Daybreak Red programme, intended for authorized offensive security testing. The company lists security requirements such as multi-factor authentication and managed devices.
What did METR conclude about Anthropic?
METR reported that Anthropic’s main progress measure grew about 1.9 times faster after October 2025 than before. It estimated AI may explain about 80% of the acceleration, while acknowledging that much of its most informative evidence was redacted.
Do the reports prove AI is causing more real-world cyberattacks?
No. The source material describes vulnerability research and test-machine exploits, but does not report that these findings caused a wave of attacks. It does suggest that cheaper vulnerability discovery could increase pressure on organisations to assess and apply patches quickly.
What should organisations and users do now?
Follow software-vendor security advisories, assess whether published issues affect deployed systems, and install relevant updates promptly. The reports do not identify a single universal patch action for all Linux users.
Source: ThorstenMeyerAI.com
