🔍 Read the full analysis: How Much Faster Is The Lab? GPT-6 Cyber And METR Offer Clues 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 November 5 and attributed a mid-October Linux security campaign to the model, reporting 948 working privilege-escalation exploits, including 588 it said were previously unknown. Separately, METR reported that Anthropic’s measured AI progress accelerated after October 2025, while warning that much of the evidence is confidential. The figures point to faster vulnerability discovery and AI-assisted research, but key claims are difficult to verify independently.
OpenAI named GPT-6 Cyber on November 5, saying the specialized model powered a mid-October campaign that produced 948 working Linux privilege-escalation exploits, 588 of them previously unknown according to the company. On the same day, independent evaluator METR reported that Anthropic’s measured rate of AI progress has accelerated, linking much of the increase to AI helping researchers build AI. Together, the reports point to faster vulnerability research and faster model development, though important parts of both accounts remain difficult to check independently.
OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra, intended for authorized security work such as finding vulnerabilities, developing exploits and testing web applications. The company says access is limited to approved organizations in its Daybreak Red programme, which covers offensive testing such as penetration testing and red-team work. Its system card lists safeguards including identity checks, hardware security keys, managed devices and restricted API access. OpenAI rates the model “Critical” under its own cybersecurity preparedness framework, the highest tier in that framework; this is a company assessment, not an independent certification.
In its account of the Linux work, OpenAI said its pipeline examined about 2.8 million lines of kernel code, or roughly 7% of the project. It reported analyzing 2,308 candidate findings, reproducing 1,724 vulnerabilities in test machines and generating 948 working local privilege-escalation exploits after grouping results by root cause. OpenAI put the API bill at $1,213,430, or about $1,279 per reported exploit. The company had described the campaign on November 4; its November 5 announcement identified GPT-6 Cyber as the model behind it.
METR’s separate report examined Anthropic’s progress over an 11-week period inside the company’s data. METR said its main measure of model capability grew about 1.9 times faster after October 2025 than before. It estimated that AI assistance may account for about 80% of the acceleration in Anthropic’s research and development. That figure is METR’s estimate, not a directly measured share of work, and the evaluator said Anthropic redacted a majority of the evidence it considered most informative.
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 Discovery Pressures Patch Teams
If OpenAI’s Linux figures are accurate, the cost of producing working exploits in this particular campaign was far below historical sums cited for individual exploits in hacking competitions. The source account compares the October result—about $1,279 per exploit—with an OpenAI run in July that reportedly found 166 Linux privilege-escalation bugs for $513,690, or about $3,100 each. These are company-reported campaigns, not a controlled independent comparison, but the reported difference illustrates why security teams are watching the cost of vulnerability research.
Lower discovery costs could benefit defenders who use automated systems to identify and fix flaws. They could also give attackers more ways to find vulnerabilities, especially if similar capabilities become broadly available. GPT-6 Cyber’s restricted access is a response to that risk, but it does not settle how quickly other models or users may acquire comparable abilities. A separate late-October Epoch AI report, as described in the source material, found that freely downloadable GLM 5.3 could turn published Firefox fixes into working attacks for about $1,000 per Firefox version. That result suggests the broader security challenge is not limited to one closed model.
For organizations, the immediate operational issue is the time between a flaw being found and a patch being installed. As vulnerability discovery becomes cheaper and potentially faster, patching capacity and response speed may matter as much as detection. For individual users, the practical implication remains straightforward: install software and security updates promptly.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition, 32GB (2024), Metallic Black
Samsung Galaxy Tab A11 (2026) – 8.7 90Hz Android Tablet, 64GB/4GB Wi‑Fi, Octa-core MediaTek Helio G99, Dual Dolby Speakers, 20W Fast Charger
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
Two Reports, Different Evidence
The OpenAI and METR developments concern different activities. GPT-6 Cyber is a model release for approved security organizations, and the Linux exploit count comes from OpenAI’s description of a specific test campaign. METR’s report instead evaluates the pace of Anthropic’s internal model development. The findings can be read as evidence of a wider pattern—AI assisting both security research and AI research—but they are not a single experiment and do not establish that every lab is advancing at the same rate.
OpenAI’s Linux numbers are company-reported. Mozilla’s public security advisories credit OpenAI with 50 high-impact Firefox fixes in October, according to the supplied source, but do not identify which model found each issue. The source also says no Linux kernel maintainer had commented publicly on the campaign at the time of writing. That leaves the public record uneven: some security work is reflected in advisories, while the model-specific contribution and full technical details are not independently confirmed there.
METR is an outside evaluator, but its report also has limits. It said Anthropic withheld much of the evidence METR found most informative, meaning readers must rely more heavily on the evaluator’s judgment than it would prefer. The source account also discloses that its original analysis was drafted with help from Anthropic’s Claude, while Anthropic and OpenAI are both subjects of the reporting. These disclosures do not establish that the findings are wrong, but they are relevant to how readers assess the analysis.
“A majority of the evidence we found most informative”
— METR, on the withheld material
What the Public Cannot Verify
The 948-exploit total and the 588 previously unknown flaws are OpenAI’s figures. The source material does not provide an independent audit of the full Linux campaign, its methodology or every reported finding. It also does not establish which model found each flaw, or whether all findings were accepted and addressed by maintainers. Mozilla’s advisories document fixes credited to OpenAI but do not attribute them to GPT-6 Cyber specifically.
METR’s estimated 1.9-times acceleration and roughly 80% AI contribution also require caution. The report is an independent evaluation, but a majority of the evidence METR found most informative was redacted, according to the evaluator. The public cannot fully reproduce the analysis from the described record. It is also unclear how far the reported pattern applies beyond Anthropic or how much of the improvement reflects AI assistance rather than other changes in research methods, staffing or computing resources.
The supplied reporting says the Linux campaign occurred in mid-October and that the November 5 announcement identified GPT-6 Cyber as its model. It does not provide a public maintainer response or a complete account of the vulnerabilities’ disclosure and remediation status. Those gaps limit conclusions about the campaign’s real-world impact.
Access, Disclosure and Follow-Up
OpenAI says GPT-6 Cyber is available through its API and Codex only to organizations it approves for Daybreak Red. The source material, citing CSO Online and The Register, says the company expected to extend access to all Red customers over the coming weeks. The next useful indicators will be whether OpenAI publishes more technical detail, whether maintainers or other independent researchers validate the Linux findings, and how many of the reported flaws have been fixed.
For METR’s findings, further reporting may clarify what evidence Anthropic withheld and whether later evaluations reproduce the reported acceleration. Until then, the 1.9-times measure and the estimate that AI accounts for about 80% of the change should be treated as METR’s conclusions with significant limits on public verification.
Security teams do not need to wait for those questions to be resolved before acting. They can keep software inventories current, prioritize vendor security updates and reduce delays between patch availability and installation. The reports do not show that every organization faces an immediate new exploit, but they do reinforce the value of rapid, routine patching as automated vulnerability discovery advances.
Key Questions
What is GPT-6 Cyber?
OpenAI describes it as a cyber-specialized version of GPT-6 Astra for authorized security work, including vulnerability discovery, exploit development and web application testing. Access is restricted to organizations approved for its Daybreak Red programme.
Did GPT-6 Cyber find 948 Linux vulnerabilities?
OpenAI reported that its campaign produced 948 working local privilege-escalation exploits, including 588 it said were previously unknown. The figures are OpenAI’s account; the supplied material does not establish that the full total has been independently audited.
What did METR conclude about Anthropic?
METR reported that its main measure of Anthropic’s AI progress grew about 1.9 times faster after October 2025 than before. It estimated that AI assistance may explain about 80% of the acceleration, while noting that much of the evidence was redacted.
Does this mean AI can already exploit any computer?
No. The reports describe particular research and testing results, not a finding that AI can exploit every system. GPT-6 Cyber is restricted, and the public record does not independently verify every reported Linux result or establish its status across real-world systems.
What should users and companies do now?
Install security updates promptly and maintain a process for prioritizing and applying patches. The reports do not identify a specific new action for every user, but they highlight why delays in fixing known flaws can matter.
Source: ThorstenMeyerAI.com
