🔍 Read the full analysis: What GPT-6 Cyber And METR Add To The AI Capability Debate 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 helped produce 948 working Linux privilege-escalation exploits in an October test, including 588 it describes as previously unknown. Separately, METR reports that Anthropic’s measured model-improvement pace increased after October 2025, but says much of the underlying evidence is redacted. Both findings point to a faster-moving AI capability debate, though key claims cannot be independently checked from public material.
OpenAI announced GPT-6 Cyber on November 5, describing it as a cyber-specialized model available only to approved security organizations, while evaluator METR published a report on Anthropic’s model progress. OpenAI says an October Linux-kernel test produced 948 working privilege-escalation exploits, 588 of them previously unknown; METR says Anthropic’s measured improvement pace has risen. The claims add to debate about how quickly AI capabilities are advancing, but much of the supporting evidence remains private.
OpenAI describes GPT-6 Cyber as a version of GPT-6 Astra designed for authorized security work, including vulnerability discovery, exploit development and web-application testing. Access is restricted to organizations approved for the company’s Daybreak Red programme, which is intended for offensive security testing such as penetration tests and red teaming. OpenAI’s system card lists requirements including identity checks, hardware security keys, managed devices and restricted API access. The model carries OpenAI’s own “Critical” cybersecurity rating, the highest tier in its Preparedness Framework; this is the company’s assessment, not an external certification.
OpenAI’s reported Linux test began with analysis of about 2.8 million lines of kernel code. The company says its process examined 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and produced 948 working local privilege-escalation exploits. It says 588 were previously unknown. These are company-reported results, and the public material does not provide an independently audited account of each finding. OpenAI reported an API cost of about $1.21 million for the run, or roughly $1,279 per working exploit.
The second development is METR’s assessment of Anthropic’s progress. After examining data from the company over 11 weeks, METR concluded that its main measure of model improvement grew about 1.9 times faster after October 2025 than before. METR estimates that AI systems themselves may account for about 80% of the acceleration by helping Anthropic’s researchers. That estimate is an interpretation of the evidence, not a direct measurement of the share attributable to AI.
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 Discovery, Faster Development
Together, the announcements concern two parts of the capability debate: AI systems that can assist with finding and exploiting software flaws, and AI systems that may help researchers produce more capable models. If the reported security results hold up, the cost of identifying vulnerabilities could fall, increasing the volume of flaws defenders need to triage and fix. That could benefit security teams using AI for authorized testing, while also raising the stakes if similar capabilities are misused.
The practical consequence is not that every reported exploit will work against real systems or that attacks are inevitable. OpenAI says access to GPT-6 Cyber is limited to vetted organizations, and the Linux results came from a controlled test. Still, a lower reported cost per exploit could change the economics of vulnerability research. For organizations, patching and monitoring may become more time-sensitive; for users, installing security updates promptly remains a direct way to reduce exposure.
METR’s finding matters because rapid progress in AI-assisted research could compound: better tools may help labs improve the next generation of tools. But the report’s estimate should not be treated as a forecast or a precise causal result. Its importance lies in identifying a possible feedback effect that deserves further measurement and independent scrutiny.
Popular right now
Samsung 65-Inch The Frame Pro Neo QLED 4K TV (2026) Bundle with Modern Brown Bezel, HDMI Cables, and Home Theater Guide
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
LG 55-Inch OLED evo G2 Series 4K Smart TV (OLED55G2PUA, 2022)
As an affiliate, we earn on qualifying purchases.
What the Two Reports Measure
OpenAI’s October Linux campaign follows an earlier test the source material says took place in July. In that run, OpenAI reported 166 Linux privilege-escalation bugs for $513,690, or about $3,100 per exploit. The comparison suggests a lower reported cost in October, but both sets of figures come from OpenAI and are not independent evaluations. The source also says Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October; those advisories do not identify which model found each flaw.
These figures measure different things. The Linux results concern exploits generated in a controlled environment, while METR’s report concerns a trend in Anthropic’s model-improvement measure. Neither, on its own, establishes how often these capabilities will translate into real-world attacks or how quickly a particular future model will improve. The public record described in the source includes company announcements, system-card material, public advisories and METR’s analysis, but not all underlying data.
““A majority of the evidence we found most informative” was redacted.”
— METR, on evidence access
Evidence Still Behind Closed Doors
OpenAI’s exploit totals and costs are based on its own account. The source material does not provide public, independent verification of every Linux result, and no Linux maintainer is identified as having commented publicly. The reported 588 previously unknown flaws also require careful interpretation: the source does not establish that every flaw has been independently confirmed or describe the disclosure and remediation status of each one.
Likewise, Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes but do not specify which model discovered which bug. METR says much of the evidence it found most informative was redacted, limiting outsiders’ ability to assess its analysis or the estimate that AI contributed about 80% of Anthropic’s acceleration. The report describes a measured trend; it does not establish that the pace will continue or that AI assistance alone caused it.
It is also not clear from the provided material how broadly GPT-6 Cyber will be deployed, how the reported test results compare with equivalent independent evaluations, or what safeguards will be used to measure misuse and defensive benefit over time.
Access, Disclosure and Follow-Up
OpenAI says GPT-6 Cyber will be available through its API and Codex to organizations it approves for Daybreak Red. The source material says the rollout to eligible Red customers is expected over the coming weeks, but does not give a complete schedule. Further public reporting may clarify how findings from the Linux campaign are being disclosed and whether independent researchers or maintainers validate the results.
For the Anthropic assessment, a key next step is additional evidence that can be examined outside the company and METR. Subsequent measurements could show whether the post-October 2025 pace persists and help distinguish AI assistance from other changes in research practices or resources. Until then, both developments are best read as consequential but partly company- and evaluator-reported signals, not definitive proof of a fixed trajectory.
Key Questions
What is GPT-6 Cyber?
GPT-6 Cyber is a cyber-specialized version of GPT-6 Astra that OpenAI says is intended for authorized security work. Access is limited to organizations approved for its Daybreak Red programme.
Did GPT-6 Cyber produce 948 confirmed Linux flaws?
OpenAI reports that its October test produced 948 working local privilege-escalation exploits, including 588 it describes as previously unknown. The figures are based on OpenAI’s account; the material provided does not show independent confirmation of every result.
What did METR conclude about Anthropic?
METR said Anthropic’s main model-improvement measure grew about 1.9 times faster after October 2025 than before. It estimated that AI may account for about 80% of the acceleration, while acknowledging that much of the supporting evidence is not public.
Are the two announcements independently verified?
Not fully on the evidence described here. OpenAI’s exploit totals and costs are company-reported, and METR says many of its most informative materials were redacted. Public advisories and the evaluator’s report provide some outside documentation, but important details remain unavailable.
What should organizations and users do now?
Organizations should maintain timely vulnerability triage, security testing and patching processes. For individual users, installing software and security updates promptly remains a practical precaution; the reports do not establish that a specific product or user is under attack.
Source: ThorstenMeyerAI.com
