🔍 Read the full analysis: A Faster AI Lab And A $1,279 Exploit: Reading GPT-6 Cyber And METR 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 GPT-6 Cyber powered security tests that produced 948 working Linux privilege-escalation exploits, including 588 it describes as previously unknown; the reported API cost was about $1,279 per exploit. Separately, METR says Anthropic’s measured model-improvement pace accelerated, while noting that much of its supporting evidence is redacted. Both developments point to faster AI capabilities, but key claims rely on company data or limited public evidence.
OpenAI said on November 5 that its restricted cybersecurity model, GPT-6 Cyber, powered an October test that produced 948 working Linux privilege-escalation exploits, at a reported API cost of about $1,279 per exploit. On the same day, independent evaluator METR reported that Anthropic’s measured model-improvement pace has increased, with much of the evidence it reviewed not publicly available.
OpenAI’s figures describe a test of Linux kernel code conducted in mid-October. The company says its process examined about 2.8 million lines of code, assessed 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. It reports producing 948 working local privilege-escalation exploits; 588, it says, were previously unknown. The reported total API bill was $1,213,430, or roughly $1,279 per working exploit. These are OpenAI’s results, not an independently verified count in the material available here.
OpenAI describes GPT-6 Cyber as a cybersecurity-focused variant of GPT-6 Astra, trained for vulnerability discovery, exploit development and web-application testing. Access is limited to organisations approved for the company’s Daybreak Red programme, intended for offensive security testing. The system card sets out safeguards including identity checks, hardware security keys, managed devices and limits on API access. OpenAI assigns the model its highest internal cybersecurity preparedness rating, “Critical”; that is the company’s own assessment, not an external certification.
METR’s report concerns a separate development at Anthropic, not the model used in OpenAI’s Linux test. After eleven weeks reviewing Anthropic data, METR said its main measure of model capability grew 1.9 times faster after October 2025 than before. METR estimated that AI assistance may account for about 80% of the acceleration in Anthropic’s research productivity. It also cautioned that Anthropic redacted a majority of the evidence METR found most informative, limiting what outsiders can check.
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 Tighten Patch Timelines
If OpenAI’s reported results hold up, the cost of producing working software exploits has fallen sharply compared with its own earlier test. The source material says a July run produced 166 Linux privilege-escalation bugs for $513,690, approximately $3,100 per exploit; OpenAI’s October figures imply about $1,279 each. That comparison is between two company-reported runs, and does not establish how the results would generalize to other code, systems or users.
The potential effect is not confined to defenders. More affordable discovery and exploit development could help security teams find and fix flaws, but could also lower the cost of preparing attacks. That makes the time between a vulnerability being found and software being patched more consequential. For organisations, the practical concern is whether security review and update processes can keep pace; for users, installing available updates promptly remains a straightforward protective step.
The METR findings add a longer-term concern: if AI tools help labs improve models faster, the capabilities being assessed may change more quickly. METR’s estimate is not a forecast of a particular release date or a direct measure of cybersecurity risk. It does, however, raise the stakes for independent evaluation and for clear reporting about what models can do.
Popular right now
Shark HydroVac MessMaster Wet/Dry Vacuum & Mop Combo (AW261)
Samsung QN65LS03FAFXZA 65 The Frame QLED 4K Art Mode Vision AI Smart TV (2025)
Ninja Mega Kitchen System Blender and Food Processor, 1500W, 72-oz Pitcher, 8-Cup Bowl, Black (BL770)
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The OpenAI announcement and METR report address related but distinct questions. OpenAI’s test concerns cybersecurity performance: whether agents can identify vulnerabilities and produce working exploits. METR’s analysis concerns the pace of AI progress inside Anthropic, including whether AI tools are helping researchers accelerate their work. The overlap is the possibility that AI capability and the means of building it are both advancing.
OpenAI said its Linux findings were disclosed to maintainers. The source material also reports that Mozilla public advisories credited OpenAI with 50 high-impact Firefox fixes in October, but those notices do not specify which model found each issue. A separate OpenAI claim cites 81 previously unknown high-severity Firefox flaws. The public advisory record described here does not establish that those figures refer to the same set of bugs or identify GPT-6 Cyber as their source.
The source material also cites Epoch AI research indicating that a freely downloadable model, GLM 5.3, could turn published Firefox fixes into working attacks for about $1,000 per Firefox version. That is a reported research result, not evidence that every publicly available model can reliably exploit software in real-world conditions. Together, the examples show why a model’s access restrictions matter, while also showing that restrictions on one company’s model cannot settle the broader security question.
““The public has to rely on METR’s judgment much more than would be ideal.””
— METR, on the limits of its public evidence
What the Public Can Verify
The Linux exploit count, the number of previously unknown flaws and the API bill are OpenAI’s reported figures. The source material does not provide an independent audit of the test, a public list of all findings or a Linux maintainer’s public confirmation of the count. It is also unclear how many of the reported flaws have been fixed, how the findings were grouped by root cause, and whether the measured cost includes expenses beyond API usage.
The relationship between OpenAI’s separate Firefox figures is also unclear: public Mozilla advisories reportedly credit OpenAI with 50 high-impact fixes, while OpenAI is said to have identified 81 previously unknown high-severity flaws. The available material does not reconcile those totals or establish which model found each flaw. For METR’s report, the 1.9-times pace comparison and estimated 80% contribution from AI remain difficult for outsiders to assess because much of the underlying evidence is redacted.
Access restrictions may reduce risks from GPT-6 Cyber, but their effectiveness has not been independently demonstrated in the information provided. It is not clear how quickly approved access will expand, how often OpenAI will review or revoke it, or how readily comparable capabilities can be reproduced with other models.
Disclosure, Patching and Access
The next useful signals will be public confirmations from Linux maintainers and more detail about the status of the reported vulnerabilities. Security advisories and patch releases can show how many findings were accepted and addressed, though they may not disclose the model or process behind each discovery. OpenAI has said the findings were shared with maintainers; the source material does not give a schedule for further technical disclosures.
OpenAI’s Daybreak Red programme is expected, according to coverage cited in the source material, to reach more approved customers over the coming weeks. That timing and the programme’s operating rules will matter as researchers assess whether screening, monitoring and revocation controls are working. The company’s published pricing and system card provide some detail, but do not substitute for independent evidence about real-world use.
METR’s report leaves open whether Anthropic or the evaluator will publish additional evidence that can be checked without revealing sensitive company data. For organisations and individual users, the near-term response is less uncertain: maintain patch processes, prioritize security updates and treat claims about AI-generated exploits as serious while distinguishing company-reported outcomes from independently confirmed incidents.
Key Questions
What did OpenAI say GPT-6 Cyber produced?
OpenAI says the model powered an October test that generated 948 working Linux privilege-escalation exploits, including 588 it describes as previously unknown. Those numbers have not been independently verified in the material provided.
How was the $1,279-per-exploit figure calculated?
It is the reported $1,213,430 API bill divided by 948 working exploits, rounded to about $1,279 each. The figure reflects API costs as described, not necessarily every cost of conducting the research.
Who can use GPT-6 Cyber?
OpenAI says access is limited to organisations it approves for the Daybreak Red programme, which is for offensive security testing. Its system card describes identity, device and account safeguards.
What did METR find about Anthropic?
METR said its main measure of model capability grew 1.9 times faster after October 2025 than before. It estimated AI assistance may account for about 80% of the acceleration, while warning that much of its supporting evidence was redacted.
Are the exploit and METR figures independently confirmed?
Not fully, based on the source material. OpenAI’s exploit totals are company-reported, and METR’s assessment relies partly on evidence that is not public. Mozilla advisories provide some public security records, but do not identify which model found each flaw.
Source: ThorstenMeyerAI.com
