🔍 Read the full analysis: How GPT-6 Cyber And METR Frame The Changing Pace Of AI on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
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 described as previously unknown. Separately, METR estimates Anthropic’s model-improvement pace has nearly doubled since October 2025, while warning that much of its evidence is not public. The reports point to faster AI-assisted security research and AI helping accelerate AI development, but both leave important details difficult to verify independently.
OpenAI has identified GPT-6 Cyber as the model behind an October campaign that it says generated 948 working Linux privilege-escalation exploits, while independent evaluator METR reported on November 5 that Anthropic’s model-improvement pace has accelerated. The two developments matter because they suggest AI systems are advancing both in finding software weaknesses and in helping researchers build more capable models, though much of the underlying evidence remains private.
OpenAI’s November 4 account of the Linux test says its systems scanned about 2.8 million lines of kernel code, roughly 7% of the project. The company says the process examined 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and produced 948 working local privilege-escalation exploits. OpenAI reported that 588 were previously unknown. The campaign ran in mid-October; OpenAI named GPT-6 Cyber as the model behind it on November 5.
OpenAI’s figures are company-reported results, not independently audited totals in the material available. The source account says the API bill was $1,213,430, or about $1,279 per working exploit. That is a calculation based on the reported cost and output, not a measure of the cost to find, validate and fix every vulnerability. Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes in October, but do not specify which model found each issue. No Linux maintainer comment is included in the source material.
GPT-6 Cyber is a cybersecurity-focused variant of GPT-6 Astra, according to OpenAI’s system card. The company restricts it to approved organisations in its Daybreak Red programme, which is intended for authorised offensive security testing. OpenAI rates the model “Critical,” the highest level in its own preparedness framework; that is the company’s assessment, not an external certification. The model is also priced for approved users at $20 per million input tokens and $100 per million output tokens, according to OpenAI’s developer page.
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
Exploit Costs and the Patch Window
The reported results point to a change in the economics of vulnerability research. OpenAI says an earlier July Linux run found 166 privilege-escalation bugs for $513,690, around $3,100 per exploit; its October figures work out to about $1,279 per exploit. Those numbers are not directly comparable measures of total security risk, but they indicate that the company’s reported cost per working exploit fell between the two tests.
Lower costs can help defenders find and repair flaws, but the same technical capability could also aid people seeking to exploit them. The source account cites Epoch AI research from late October saying a freely downloadable model, GLM 5.3, used published Firefox fixes to produce working attacks for about $1,000 per Firefox version. That is a separate reported demonstration, not evidence that GPT-6 Cyber was used in an attack. For organisations, the practical concern is that the time available to patch a disclosed weakness may shrink as vulnerability discovery and exploit development become cheaper.
METR’s report adds a second concern: AI may be speeding up the work that produces future AI systems. Its estimate that AI accounts for most of Anthropic’s acceleration is an interpretation of evidence gathered during an eleven-week review, not a direct measurement of every research task. If that estimate is broadly right, progress in AI capability could reinforce itself, making monitoring and reliable public evidence more important.
Products Worth Considering
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
Samsung Galaxy Tab A11 Wi‑Fi (International Version), 8.7, 64GB + 4GB, Gray
Amazon Kindle Paperwhite Signature Edition, 32GB (2024), Metallic Black
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The OpenAI and METR developments are related by theme, but they are not one joint study. OpenAI reported a model’s cybersecurity performance and then identified the model used. METR, an independent evaluator, examined Anthropic’s rate of progress using access to company data over eleven weeks. The reports concern different organisations and different measures; neither establishes that the Linux work caused or resulted from Anthropic’s reported acceleration.
METR says its main measure of Anthropic’s model improvement grew about 1.9 times faster after October 2025 than before. It estimates that roughly 80% of the acceleration may come from AI helping Anthropic’s own researchers. METR also cautioned that Anthropic withheld much of the evidence it found most informative, limiting what outsiders can check.
The security claims have their own verification limits. OpenAI’s system card and announcement provide the main detail about GPT-6 Cyber and the Linux campaign. Mozilla’s advisories offer a separate public record of Firefox fixes, but do not connect each fix to a model. The source material reports no public response from Linux maintainers, so the full disposition of the kernel findings is not established here.
“The public, METR writes, “has to rely on METR’s judgment much more than would be ideal.””
— METR
What the Public Can Verify
The source material does not establish how many of the 588 reportedly new Linux flaws have been independently confirmed, patched or assigned public identifiers. It also does not include a public assessment from Linux maintainers or a breakdown linking each confirmed vulnerability to GPT-6 Cyber. OpenAI’s total should therefore be treated as the company’s account of its test, not as a fully public tally of verified and remediated bugs.
For METR’s findings, the main uncertainty is evidentiary: the evaluator says Anthropic redacted a majority of the evidence it found most informative. METR’s 1.9-times pace estimate and its roughly 80% estimate for AI’s contribution are its conclusions from the review, but outsiders cannot inspect much of the material used to reach them. The source account also does not provide the full methodology or enough detail to independently reproduce those estimates.
It remains unclear how quickly access to GPT-6 Cyber will expand beyond approved Daybreak Red organisations, how its safeguards will perform in practice, and whether the cost figures capture all testing and validation work. The reports describe capabilities and estimates, not evidence of a specific real-world attack or proof that the same results will occur across other systems.
Access, Disclosure and Patching
OpenAI says GPT-6 Cyber is limited to approved organisations, with access expected to reach Daybreak Red customers over the coming weeks, according to reports cited in the source material. Its stated controls include identity checks, multifactor authentication, managed devices and restrictions on API access. How those controls operate across the programme, and whether OpenAI will publish additional evaluation results, remain points to watch.
The immediate security priority is the handling of reported vulnerabilities. Public advisories, maintainers’ responses and patch releases would help clarify which findings were confirmed and what users need to update. Organisations should follow vendor and project security notices and apply relevant fixes promptly; the material does not identify a specific Linux update that readers should install.
METR’s assessment also leaves a clear next step for public understanding: more disclosure about the evidence and methods behind its estimates. Until then, the reported acceleration is an informed external assessment with acknowledged limits, while OpenAI’s exploit totals remain company-reported. Further technical documentation and independent confirmation would help distinguish demonstrated results from claims that cannot yet be checked.
Key Questions
What is GPT-6 Cyber?
GPT-6 Cyber is a cybersecurity-focused variant of GPT-6 Astra, described by OpenAI as designed for authorised security work. OpenAI limits access to approved organisations in its Daybreak Red programme.
Did GPT-6 Cyber find 948 Linux vulnerabilities?
OpenAI says its October campaign produced 948 working local privilege-escalation exploits, including 588 it described as previously unknown. The source material does not provide an independent audit of those totals or a public confirmation of every finding.
What did METR find about Anthropic?
METR reported that its main measure of Anthropic’s model-improvement pace grew about 1.9 times faster after October 2025 than before. It estimated that AI may account for roughly 80% of the acceleration, while warning that much of its most informative evidence was not public.
Does the reporting show that AI was used in a real-world cyberattack?
No. The Linux figures describe a reported test campaign, and the source material does not document a real-world attack using GPT-6 Cyber. It does report a separate Epoch AI demonstration involving a downloadable model and Firefox fixes, which is not evidence of an attack by GPT-6 Cyber.
What should computer users do now?
Users should install security updates from their software and operating-system providers promptly and follow relevant security advisories. The source material does not name a particular Linux patch or update for all users to install.
Source: ThorstenMeyerAI.com
