🔍 Read the full analysis: A Closer Look At GPT-6 Cyber And METR's AI Speed Findings 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 named GPT-6 Cyber on Nov. 5, attributing a Linux security campaign that produced 948 working privilege-escalation exploits to the model. Separately, METR reported that Anthropic’s measured AI progress accelerated after October 2025, but said much of its supporting evidence is confidential. The claims point to faster vulnerability discovery and AI-assisted model development, though independent verification is limited.
OpenAI named GPT-6 Cyber on Nov. 5 as the model behind a mid-October Linux security campaign that it says produced 948 working privilege-escalation exploits, including 588 for flaws not previously known. On the same day, independent evaluator METR reported that Anthropic’s measured AI progress has accelerated, while warning that much of its most informative evidence is not public. Together, the announcements describe two consequential but differently supported developments: AI systems finding software vulnerabilities and AI potentially helping labs improve their models faster.
OpenAI describes GPT-6 Cyber as a cyber-focused version of GPT-6 Astra, built for authorized security work such as vulnerability discovery, exploit development and web application testing. Access is restricted to organisations approved for the company’s Daybreak Red programme, which covers offensive security testing. The system card describes controls including identity checks, hardware security keys, managed devices and limits on approved IP addresses. OpenAI says it can monitor, pause or revoke access. The company’s own Preparedness Framework rates the model “Critical” for cybersecurity, its highest tier; that rating is an internal assessment, not an external certification.
OpenAI’s account of the Linux work says its pipeline examined about 2.8 million lines of kernel code, identified 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. It then produced 948 working local privilege-escalation exploits, grouping findings by root cause. OpenAI put the API cost at $1,213,430, or roughly $1,279 per working exploit. Those figures come from the company; the available material does not establish that an independent party has reproduced the campaign or verified every count.
METR’s separate report says its measure of Anthropic’s AI progress rose about 1.9 times faster after October 2025 than it had before. METR estimates that AI assistance to Anthropic’s researchers accounts for roughly 80% of the acceleration. It also says Anthropic redacted a majority of the evidence METR found most informative, limiting what outsiders can independently inspect. The report is an evaluator’s assessment of evidence shared by Anthropic, not a public audit of the company’s internal 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
Faster Discovery, Shorter Patch Windows
If OpenAI’s Linux figures hold up, the campaign illustrates how AI tools can reduce the time and expense needed to find and turn software flaws into working exploits. That matters to both sides of cybersecurity: defenders may use similar systems to locate and repair vulnerabilities, while attackers could benefit if comparable capabilities become accessible outside controlled programmes. The reported cost of about $1,279 per exploit is a figure for this particular OpenAI run, not a universal market price, but it highlights why security teams may need to treat vulnerability discovery as increasingly scalable.
The METR findings add a related concern for oversight. If AI is helping researchers build or improve later AI systems, advances may compound faster than external observers can evaluate them. METR’s estimate is uncertain and rests partly on non-public evidence, so it does not establish a precise rate of future progress. Still, the report makes transparency and independent evaluation more relevant as labs deploy increasingly capable models.
For organisations, the immediate practical issue is response time. Finding flaws more cheaply does not automatically make systems safer: security teams still need to validate reports, develop fixes and get patches installed. For individuals, the source material’s grounded takeaway is straightforward: apply software updates promptly, especially for browsers and operating systems.
Popular right now
Ninja Foodi DZ201 8-Quart DualZone 6-in-1 Air Fryer
Samsung 65-Inch The Frame Pro Neo QLED 4K TV (2026) Bundle with Modern Brown Bezel, HDMI Cables, and Home Theater Guide
Ninja Flip Toaster Oven & Air Fryer, 8-in-1, Stainless Steel, SP151
As an affiliate, we earn on qualifying purchases.
The Security Work Behind the Claims
The Linux campaign was described by OpenAI on Nov. 4, before the company identified GPT-6 Cyber as the model behind it the next day. OpenAI said the work ran in mid-October. Its account gives several stages of analysis, from candidate findings to reproduced vulnerabilities and working exploits; those stages are distinct, and the 948 figure refers to working local privilege-escalation exploits rather than every candidate or reproduced flaw.
OpenAI also cited Firefox security work. The source material reports that Mozilla’s public advisories credited OpenAI with 50 high-impact Firefox fixes in October. Those advisories do not identify which model found each flaw, so they cannot independently confirm that GPT-6 Cyber was responsible for all or any particular fixes. The source also reports that OpenAI previously described a July Linux run yielding 166 privilege-escalation bugs at a stated cost of $513,690, compared with the October campaign’s reported 948 exploits and $1.21 million cost. These are company-reported results from two runs, not a controlled independent comparison.
Other research described in the source suggests that offensive techniques may not remain limited to restricted services. Epoch AI reported in late October that a downloadable model called GLM 5.3 used published Firefox fixes to produce working attacks for about $1,000 per Firefox version. That is a separate finding, with its own methodology and scope; it does not establish that the model can discover new flaws in the same way as OpenAI’s reported campaign. The distinction matters when comparing costs, capabilities and access.
““the public has to rely on METR’s judgment much more than would be ideal.””
— METR, discussing the limits of its evidence
What the Public Cannot Verify
The exploit counts, novelty count, scanning scope and cost are OpenAI’s reported figures. The supplied material does not describe an independent replication of the Linux campaign, and it says no Linux maintainer had publicly commented on the results at the time of publication. It is also unclear from the public account how the 588 previously unknown flaws were classified, how many have been fixed, or whether all affected systems have received patches.
Mozilla’s advisories document credited fixes but do not connect specific vulnerabilities to a particular model. The METR report, meanwhile, is constrained by confidential Anthropic evidence; its 1.9-times pace comparison and estimate that AI accounts for about 80% of acceleration should be understood as METR’s conclusions, not fully reproducible public measurements. The source material does not provide enough detail to resolve the effects of redactions on those estimates.
There is also no basis here to conclude that the reported costs or results will generalize to other software, models or operators. GPT-6 Cyber’s restricted availability may limit immediate access, but the separate GLM 5.3 account suggests that some security-relevant capabilities may exist in downloadable systems. Their practical reach and reliability remain unsettled.
Access, Fixes and Further Evidence
OpenAI’s stated next step is to make GPT-6 Cyber available to approved Daybreak Red organisations; the source material, citing technology news outlets, says access was expected to expand over the following weeks. OpenAI has not provided, in the material here, a public timetable for broader availability. Security teams and researchers will be watching for further technical details, public advisories and evidence that lets outside experts assess the Linux findings.
For affected software, the key next developments are confirmation of fixes and whether maintainers publish additional information about vulnerabilities and remediation. The supplied source does not establish a complete Linux patch status. Users and organisations should follow official update notices and apply security patches as they become available, while companies should review how quickly they can triage and deploy fixes.
METR’s findings may also be revisited as more evidence becomes available or as evaluators publish further assessments. For now, its report supports a cautious conclusion: Anthropic’s measured pace increased, and METR believes AI assistance explains much of that change, but the private evidence limits independent checking. Both stories will become clearer through public technical records, independent replication and later assessments.
Key Questions
What is GPT-6 Cyber?
OpenAI describes it as a cyber-focused version of GPT-6 Astra for authorized security tasks, including vulnerability discovery, exploit development and web testing. Access is restricted to organisations approved for its Daybreak Red programme.
Did OpenAI independently verify the 948 Linux exploits?
The source material presents the 948-exploit figure as OpenAI’s account. It does not describe an independent replication or public confirmation by Linux maintainers.
What did METR report about Anthropic?
METR said its measure of Anthropic’s AI progress grew about 1.9 times faster after October 2025 than before, and estimated that AI assistance accounts for roughly 80% of the acceleration. METR also said much of its most informative evidence was redacted.
Does the report show that users are already exposed to these exploits?
No. The reported campaign describes exploits produced in testing, but the supplied material does not establish that they were used against real-world systems or that particular users were affected. The public status of fixes for all reported Linux flaws is also unclear.
What should computer users do now?
Install security updates promptly and follow notices from operating-system and browser providers. The reports do not identify a specific action users must take beyond keeping software updated.
Source: ThorstenMeyerAI.com
