🔍 Read the full analysis: Could AI Labs Be Running Twice As Fast? What GPT-6 Cyber And METR Show on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI says GPT-6 Cyber helped produce 948 working Linux privilege-escalation exploits in an October test, including 588 it describes as previously unknown. Separately, METR estimates Anthropic’s model-development pace has accelerated to about 1.9 times its earlier rate, with AI contributing much of the gain. Both findings depend partly on evidence that is not public.
OpenAI named GPT-6 Cyber on November 5, saying the cybersecurity-focused model powered an October test that generated 948 working Linux privilege-escalation exploits, including 588 flaws the company says were previously unknown. The same day, independent evaluator METR reported that Anthropic’s measured model-development pace rose to about 1.9 times its earlier rate, with AI likely accounting for much of that acceleration. The announcements point to faster AI-assisted security research and AI development, but key evidence remains private or has not been independently verified.
OpenAI’s account of the Linux campaign says its system scanned about 2.8 million lines of kernel code, identified 2,308 candidate findings and reproduced 1,724 vulnerabilities in test machines. The resulting 948 working exploits were local privilege-escalation techniques, which could let a user with ordinary access gain administrator privileges. OpenAI said 588 were previously unknown. The company reported an API cost of $1,213,430, or about $1,279 per exploit, based on its figures.
OpenAI describes GPT-6 Cyber as a specialized version of GPT-6 Astra for authorized security work, including vulnerability discovery, exploit development and web-application testing. Access is limited to approved organizations in its Daybreak Red programme, intended for offensive security testing. OpenAI assigns the model a “Critical” cybersecurity rating under its own Preparedness Framework; that rating is the company’s assessment, not an independent evaluation. Its listed price for approved organizations is $20 per million input tokens and $100 per million output tokens.
METR’s separate analysis concerns Anthropic’s development process, not GPT-6 Cyber. METR says its main measure of Anthropic’s progress grew 1.9 times faster after October 2025 than before. The evaluator estimates that AI may account for about 80% of the acceleration by helping Anthropic researchers do their work. That figure is an estimate, not a directly measured share of all research output.
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 Shrink Patch Windows
If OpenAI’s reported results are representative, AI tools can reduce the cost and time needed to find and turn software flaws into working exploits. That could help defenders audit code and reproduce bugs, but the same capability can benefit people seeking to exploit vulnerable systems. The cost figure is a company-reported API bill, not a full accounting of staff, computing or verification costs, and should not be read as the total cost of producing a reliable exploit.
For software makers and organizations, the practical concern is the time between a flaw becoming discoverable and a fix reaching users. Automated discovery can increase the volume of issues that security teams need to review, making triage, patching and deployment speed more consequential. For individual users, keeping software updated remains a useful safeguard, though the reports do not establish that a particular user or system faces an immediate new threat.
METR’s findings add a broader concern: if AI is helping labs improve AI systems faster, model capabilities may advance more quickly than expected from earlier trends. That possibility matters for security planning and oversight. But METR’s estimate applies to Anthropic’s observed development pace and should not be generalized as a measured acceleration across every AI lab.
Products Worth Considering
DJI Osmo Mobile 7P Smartphone Gimbal Stabilizer with Mic Mini Transmitter
LG UltraGear 32GS60QC-B 32-inch Curved QHD Gaming Monitor (180Hz, 1ms, 1000R, AMD FreeSync, HDR10)
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The developments are related by a possible feedback loop—AI systems helping identify software weaknesses while also helping researchers build more capable systems—but they are separate findings with different methods. OpenAI’s Linux figures come from the company’s account of a controlled campaign. METR’s analysis examines data from Anthropic and estimates the contribution of AI to the lab’s faster progress.
OpenAI also reported that an earlier Linux run in July found 166 privilege-escalation bugs at a stated API cost of $513,690, about $3,100 per exploit. The company’s October figure is lower per exploit, but the comparison is based on OpenAI’s reported counts and costs; the source material does not provide a complete independent audit of either run. Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes in October, but do not identify which model found each flaw. A separate Epoch AI test reportedly found that the downloadable model GLM 5.3 could turn published Firefox fixes into working attacks for about $1,000 per Firefox version, illustrating that advanced security capabilities are not limited to closed commercial systems.
METR says Anthropic redacted a majority of the evidence it considered most informative. It cautions that readers consequently must rely more heavily than ideal on METR’s judgment. The reports therefore offer signals about capability and pace, not a full public record that allows every underlying result to be reproduced.
“A majority of the evidence we found most informative”
— METR, on its access to Anthropic evidence
Independent Checks Remain Limited
The Linux exploit totals, the count of previously unknown flaws and the API spending figure are OpenAI’s reported numbers. The source material does not include a public, independent technical audit of the full campaign. No Linux maintainer had commented publicly at the time covered by the report, so it is unclear how many findings have been accepted, fixed or disclosed through the kernel process.
Mozilla’s advisories confirm that OpenAI received credit for 50 high-impact Firefox fixes in October, but they do not establish which model discovered those vulnerabilities or connect them to the Linux campaign. The supplied evidence also does not show whether the reported exploits work outside the test environments, how many have been patched, or whether they have been used against real systems.
METR’s 1.9-times figure depends on its chosen measure and access to Anthropic data; much of the supporting evidence is not public. Its estimate that AI accounts for about 80% of the acceleration is an interpretation of that evidence. It remains unclear whether the same pattern holds at other labs or will persist over time.
Disclosure and Follow-Up Matter
OpenAI says GPT-6 Cyber is restricted to organizations it approves for Daybreak Red, with access expected to expand to eligible Red customers over the coming weeks, according to reports cited in the source material. The company says it can monitor usage and pause or revoke access. The next useful checks will be whether OpenAI publishes additional methodology and whether Linux maintainers or security advisories document the handling and remediation of the reported findings.
Security teams will need to track patches and disclosures tied to the reported flaws rather than assume that every reproduced finding affects deployed systems. METR’s findings may also prompt further scrutiny of how AI research tools affect lab productivity. More public evidence from Anthropic or independent replication would help establish whether the estimated acceleration is robust. Until then, the two reports support a cautious conclusion: AI-assisted security work and model development may be speeding up, but the scale and wider effects are not yet fully verified.
Key Questions
What is GPT-6 Cyber?
OpenAI describes GPT-6 Cyber as a cybersecurity-focused version of GPT-6 Astra for authorized work such as finding vulnerabilities and testing applications. Access is limited to organizations approved for its Daybreak Red programme.
Did GPT-6 Cyber find 948 new Linux vulnerabilities?
OpenAI reported 948 working local privilege-escalation exploits and said 588 were previously unknown. The figures come from the company’s account, and the supplied material does not establish that all 948 were new vulnerabilities or independently verified.
What did METR find about Anthropic?
METR said its main measure of Anthropic’s model-development progress grew about 1.9 times faster after October 2025 than before. It estimated that AI contributed about 80% of the acceleration, while noting that much of the evidence it considered informative was redacted.
Does the reporting show that AI exploits are being used in real attacks?
No. The Linux campaign described was conducted in test machines, and the supplied material does not establish that its reported exploits were used against real systems. It also does not specify the current patch status of the findings.
Source: ThorstenMeyerAI.com
