GPT-6 Cyber’s Exploit And METR’s Lab: A Look At AI’s Pace

🔍 Read the full analysis: GPT-6 Cyber's Exploit And METR's Lab: A Look At AI's Pace on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

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, including 588 flaws the company says were previously unknown. On the same day, independent evaluator METR reported that a measure of Anthropic’s model-development progress grew 1.9 times faster after October 2025, linking much of the acceleration to AI assistance. The results point to two related changes—AI systems finding software weaknesses and helping researchers improve AI—but key evidence remains controlled by the companies or redacted.

OpenAI described the Linux work in a November 4 post, before naming GPT-6 Cyber as the model behind it the following day. According to figures attributed to OpenAI, the pipeline examined about 2.8 million lines of Linux kernel code, assessed 2,308 candidate findings, reproduced 1,724 vulnerabilities in test machines and yielded 948 working local privilege-escalation exploits. OpenAI says 588 of the final exploits concerned previously unknown flaws. These are company-reported results; the source material does not provide an independent validation of the full set.

The reported API bill was $1,213,430, or about $1,279 per working exploit. For comparison, an earlier OpenAI Linux run in July reportedly produced 166 exploits for $513,690, about $3,100 each. Those figures suggest a lower reported cost per exploit between the two runs, but they do not establish that the results were equivalent in difficulty, coverage or independent verification. OpenAI also reported finding 81 previously unknown high-severity Firefox flaws. Mozilla’s public advisories credit OpenAI with 50 high-impact Firefox fixes in October, but do not identify which model found each issue.

GPT-6 Cyber is described by OpenAI as a cybersecurity-focused variant of GPT-6 Astra, intended for authorized security work. Access is limited to organisations approved for the company’s Daybreak Red programme, which covers offensive testing such as penetration tests and red-team work. OpenAI rates the model “Critical,” the highest tier in its own cybersecurity preparedness framework; that rating is an internal assessment, not an independent certification.

METR’s separate report examined Anthropic’s progress over 11 weeks. It says its main measure of Anthropic’s model-development pace grew 1.9 times faster after October 2025 than before. METR estimates that AI assistance accounts for about 80% of the acceleration, while warning that Anthropic redacted a majority of the evidence METR found most informative. The estimate is therefore an evaluator’s judgment based on a limited public record, not a fully reproducible public analysis.

Cheaper Exploit Discovery Tests Defenses

If OpenAI’s reported results hold up, the cost of generating working software exploits is falling, which could change the workload for security teams. A tool that identifies weaknesses more quickly can help defenders find and fix bugs, but the same capability can increase risk if used outside authorized testing. The central practical concern is the time between a flaw being discovered and a fix reaching affected systems.

The reported Linux results do not show that the exploits were deployed against real systems, and the source material does not describe a public incident connected to the test. OpenAI says its programme restricts access and monitors use. Still, those controls apply to its own service; they do not prevent other developers from producing capable tools or attackers from using different methods.

METR’s findings matter for a related reason: if AI tools are helping researchers build stronger AI systems, development could become faster through a feedback loop. METR’s estimate does not prove that this trend will continue at the same rate, or that other labs will see the same effect. Taken together, the reports highlight both sides of accelerating AI capability: more capacity for security research and a potentially faster-moving threat environment.

Products Worth Considering

Bose TV Speaker (2nd Gen) Compact Soundbar

AmazonView Latest Price

DJI Osmo Mobile 8 3-Axis Phone Gimbal Stabilizer

AmazonView Latest Price

Apple Watch SE 3 GPS Smartwatch, 40mm

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

What the Two Reports Measure

The OpenAI announcement combines a model release with results from security testing conducted in mid-October. GPT-6 Cyber is available only to organisations OpenAI approves for the Daybreak Red tier, according to the supplied material. OpenAI lists technical access requirements including single sign-on, multifactor authentication, managed devices and restricted API access. Its safeguards and the model’s “Critical” classification are company descriptions, not outside findings.

Some outside evidence offers a partial check on the vulnerability claims. Mozilla’s October advisories credit OpenAI with 50 high-impact Firefox fixes, while not linking individual fixes to GPT-6 Cyber. The source material says no Linux maintainer had commented publicly on OpenAI’s reported Linux findings at the time it was written. Separately, Epoch AI reported in late October 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 separate evaluation and should not be treated as verification of OpenAI’s Linux results.

METR is an independent evaluator, but its Anthropic assessment has an important evidence limitation: the company withheld much of the material METR considered most informative. The report’s central pace estimate is not directly comparable to OpenAI’s exploit counts. One concerns a measure of a lab’s model-development progress; the other concerns cybersecurity tasks reported from a particular test.

“The public has to rely on METR’s judgment much more than would be ideal.”

— METR, on the limits of the public evidence

How Much Evidence Is Public

The full technical record behind OpenAI’s Linux totals is not included in the supplied material, and it is unclear how many findings have been independently reproduced or reviewed by Linux maintainers. OpenAI’s distinction between previously unknown flaws and other findings also needs to be read as the company’s own accounting. Mozilla’s public advisories provide evidence of Firefox fixes credited to OpenAI, but do not establish which model identified them.

METR’s 1.9-times pace estimate and its attribution of about 80% of acceleration to AI assistance remain estimates. Because Anthropic redacted much of the evidence, readers cannot inspect the full basis for the conclusions. The source material also does not establish whether the observed acceleration will persist, or whether it reflects changes that apply across other AI labs.

It is also unclear how comparable the July and October exploit runs were, beyond their reported counts and costs. The source material does not show how the results map onto real-world attacks, how quickly affected Linux flaws will be patched, or whether access controls around GPT-6 Cyber have been independently audited.

Patch Response and Independent Checks

The immediate next step is for maintainers and security teams to review reported findings, confirm which systems are affected and publish fixes or advisories where needed. The supplied material does not give a timetable for Linux responses or say which of the reported findings have already been patched. Organisations should continue applying security updates promptly and follow advisories from the software projects they use.

OpenAI’s Daybreak Red access is expected, according to the source material’s account of reporting by CSO Online and The Register, to reach approved Red customers over the coming weeks. OpenAI says it can monitor use and suspend access, but further details on programme oversight or independent testing were not provided here.

For METR’s assessment, greater access to supporting evidence would allow outside researchers to scrutinize the pace estimate and its attribution to AI-assisted work. Until then, the report is best read as a documented but partly opaque evaluation. Further public responses from Anthropic, OpenAI, Mozilla or Linux maintainers could clarify how much of the reported capability has been verified and what changes defenders should prioritize.

Key Questions

What is GPT-6 Cyber?

OpenAI describes it as a cybersecurity-focused variant of GPT-6 Astra for authorized security work. Access is restricted to organisations it approves for the Daybreak Red programme.

Did GPT-6 Cyber produce 948 verified Linux exploits?

OpenAI says the October test produced 948 working exploits, including 588 it classifies as previously unknown. The supplied material does not include an independent validation of all the findings, and it says Linux maintainers had not publicly commented at the time.

What did METR find about Anthropic?

METR reported that its measure of Anthropic’s model-development progress grew 1.9 times faster after October 2025 than before. It estimates AI assistance accounts for about 80% of the acceleration, while noting that much of the relevant evidence was redacted.

Does the reported cost mean exploit discovery is now cheap for everyone?

No. OpenAI reported an average of about $1,279 per working exploit for its October run, but that figure is based on a specific company test and does not establish costs for other users, models or kinds of vulnerabilities.

What should computer users do now?

Install operating-system and software security updates promptly, and follow advisories from the projects and vendors whose products they use. The reports do not identify a new public attack requiring a different immediate response.

Source: ThorstenMeyerAI.com