🔍 Read the full analysis: Inside The GPT-6 Cyber Findings And METR Report On Faster AI Labs on ThorstenMeyerAI.com
Get hardware and tech essentials 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 GPT-6 Cyber helped find and produce 948 working Linux privilege-escalation exploits, including 588 it says were previously unknown; access is limited to approved security organisations. Separately, METR reports that Anthropic’s model-improvement pace accelerated after October 2025, and estimates AI may account for much of that change. Important evidence behind both developments is not fully public.
OpenAI announced GPT-6 Cyber on November 5, describing it as a restricted model for authorized security work and linking it to a Linux campaign that produced 948 working privilege-escalation exploits. On the same day, evaluator METR published a report finding that Anthropic’s model-improvement pace accelerated after October 2025, with AI itself likely contributing substantially. The findings point to two related developments: AI systems are being used to find and exploit software flaws, while AI may also be helping labs develop models faster.
OpenAI says its agents examined about 2.8 million lines of Linux kernel code, identified 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. The company reports that the work yielded 948 working local privilege-escalation exploits, including 588 flaws it says were previously unknown. The run occurred in mid-October; OpenAI later said GPT-6 Cyber powered it. These figures come from OpenAI’s account, and the source material does not include an independent audit of the campaign.
OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra, trained for tasks including finding vulnerabilities, building exploits and testing web applications. Access is limited to organisations approved for the Daybreak Red programme, which covers offensive security testing. Its requirements include identity checks for researchers, managed devices, multifactor authentication and restricted API access. OpenAI lists a price of $20 per million input tokens and $100 per million output tokens for approved organisations.
METR’s separate report examined Anthropic data over eleven weeks. It found that its main measure of model improvement grew 1.9 times faster after October 2025 than before. METR estimates that AI assistance may explain about 80% of the acceleration. That figure is an estimate, not a directly verified measure of what caused the change; METR says much of the relevant evidence is not available publicly.
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
Security Teams Face Shorter Patch Windows
If AI can identify vulnerabilities and help turn them into working exploits more cheaply, organisations may have less time between a flaw being discovered and an attack becoming practical. OpenAI’s reported Linux results suggest substantial capability in a controlled research setting, but they do not establish how often the model can find exploitable flaws in real-world systems or how the results would translate to attackers.
The reported cost was about $1,279 per working Linux exploit, based on OpenAI’s stated API bill and exploit count. The source compares this with a July OpenAI run that produced 166 exploits for $513,690, or roughly $3,100 each. These are company-reported results from separate runs, not a general market price or a guarantee of repeatable performance. The broader concern is that defenders may need to triage and patch vulnerabilities faster as automated discovery becomes more accessible.
METR’s findings matter for a different reason: if AI tools are helping researchers improve AI systems more quickly, the pace of capability development could itself accelerate. The report does not settle how much AI caused the change, but it makes the question relevant to oversight, safety planning and the resources available to evaluate new models.
Products Worth Considering
Ledger Flex + Folio Crypto Hardware Wallet
Trezor Safe 5 Crypto Hardware Wallet with Color Touchscreen, Black Graphite
Trezor Safe 7 Crypto Hardware Wallet
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The Linux findings and METR report are separate developments, not one coordinated study. OpenAI’s claims concern a cybersecurity model and a specific campaign against Linux kernel code. METR evaluated the pace of progress at Anthropic using data from inside the company. Their publication on the same day connects them as evidence of a wider trend, but neither report independently verifies the other.
OpenAI says its reported Linux work followed a July run that found 166 privilege-escalation bugs. Mozilla’s public advisories, according to the source material, credit OpenAI with 50 high-impact Firefox fixes in October, but do not identify which model found each flaw. The source also cites Epoch AI’s late-October demonstration that a downloadable model, GLM 5.3, could turn published Firefox fixes into working attacks for about $1,000 per Firefox version. That example concerns reproducing attacks from disclosed fixes, not discovering the same vulnerabilities from scratch.
OpenAI rates GPT-6 Cyber “Critical” under its own Preparedness Framework, its highest cybersecurity tier. That is OpenAI’s internal assessment, not an independent certification. METR, meanwhile, says Anthropic withheld much of the evidence it considered most informative, limiting what outside readers can check about its conclusions.
“the public has to rely on METR’s judgment much more than would be ideal”
— METR, in its report on Anthropic
What the Public Cannot Verify
The Linux totals, including the number of previously unknown flaws and the API bill, are OpenAI’s figures. The available material does not provide a public, independent review of the full dataset, the method used to classify findings as new, or the exploit validation process. The account also does not establish whether all 948 exploits were shared with maintainers or how many flaws have been fixed.
The source material says no Linux maintainer had commented publicly at the time of writing. It also does not identify which model discovered each Firefox issue credited to OpenAI by Mozilla. OpenAI’s access controls and monitoring are described by the company; their effectiveness in preventing misuse has not been independently demonstrated in the material provided.
METR’s estimate that AI accounts for about 80% of Anthropic’s faster progress is likewise uncertain. The report’s underlying evidence is partly confidential, and its estimate should not be read as a direct measurement of AI’s causal contribution. It remains unclear how durable the acceleration will be and whether the same pattern applies across other labs.
Disclosure, Access and Follow-Up
OpenAI says GPT-6 Cyber is available only to approved Daybreak Red organisations, with broader access for eligible Red customers expected over the coming weeks, according to reports cited in the source material. OpenAI can pause or revoke access under its stated programme rules. The company’s next public disclosures, if any, may clarify how findings were validated, reported to maintainers and addressed.
For Linux users and organisations, the immediate practical step remains applying security updates promptly. The supplied material does not list specific affected kernel versions or confirmed patch status, so readers should follow their Linux distribution’s security advisories rather than infer that a particular system is vulnerable.
Further public information is also needed to assess METR’s conclusion, including whether Anthropic or METR releases more detail about the withheld evidence and how the measure is calculated. Until then, the report is evidence of a reported acceleration, not a definitive account of its cause. Both developments warrant scrutiny as more technical documentation and maintainer responses become available.
Key Questions
What did OpenAI announce?
OpenAI announced GPT-6 Cyber, a cyber-focused model it says is intended for authorized security work. Access is restricted to organisations it approves through its Daybreak Red programme.
Did GPT-6 Cyber find 948 Linux vulnerabilities?
OpenAI says its mid-October campaign produced 948 working local privilege-escalation exploits and that 588 involved previously unknown flaws. Those totals are company-reported; the source material does not provide an independent audit.
What did METR conclude about Anthropic?
METR reported that its main measure of Anthropic’s model improvement grew 1.9 times faster after October 2025 than before. It estimated that AI may account for about 80% of the acceleration, while acknowledging limits in the public evidence.
Does the Linux report show that ordinary users are at risk now?
Not by itself. The source material does not identify affected versions, confirm which reported flaws remain unpatched, or describe a live attack. Users and organisations should follow official security advisories and install relevant updates promptly.
What remains unverified?
Independent confirmation of OpenAI’s exploit totals and costs is not provided, and the public cannot inspect much of the evidence behind METR’s estimate. Maintainer responses, patch status and the precise causal role of AI in Anthropic’s faster progress remain unclear.
Source: ThorstenMeyerAI.com
