Reading The GPT-6 Cyber And METR Findings On AI Capability Growth

🔍 Read the full analysis: Reading The GPT-6 Cyber And METR Findings On AI Capability Growth on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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, was behind tests that produced 948 working Linux privilege-escalation exploits, including 588 that the company says were previously unknown. On the same day, independent evaluator METR reported that Anthropic’s measured AI capabilities are improving faster than before, while saying much of the evidence behind its assessment remains confidential. The reports point to two related developments: AI systems are being used to find software vulnerabilities, and AI may be helping accelerate the work of building more capable systems.

OpenAI describes GPT-6 Cyber as a cybersecurity-focused version of GPT-6 Astra, intended for authorized security work such as vulnerability discovery, exploit development and application testing. Access is limited to organizations approved for the company’s Daybreak Red programme, which is designed for offensive security testing. OpenAI’s system card says participating organizations face controls including multifactor authentication, managed devices and restricted API access. These are company-reported safeguards; their effectiveness has not been independently established in the material available.

OpenAI’s account of the Linux test says its process examined about 2.8 million lines of kernel code, generated and reviewed candidate findings, reproduced vulnerabilities in test machines and produced 948 working local privilege-escalation exploits. The company says 588 were previously unknown. It reported an API cost of about $1.21 million, or roughly $1,279 per working exploit. These figures describe OpenAI’s own test and accounting; the source material does not include an independent audit of the results or cost calculation. Linux maintainers had not publicly commented in the cited material.

METR’s report, published November 5 after an eleven-week evaluation using Anthropic data, says its main measure of Anthropic’s AI capability growth increased about 1.9 times faster after October 2025 than before. METR estimates that AI assistance may explain about 80% of the acceleration. That estimate is an interpretation of the evidence, not a directly measured share of Anthropic’s research work. METR also says Anthropic withheld a majority of the evidence it considered most informative, limiting how fully outsiders can assess the conclusion.

Faster Discovery Tightens Patch Timelines

The reports matter because vulnerability discovery and software repair are linked but separate tasks. If AI tools make it faster or cheaper to identify exploitable flaws, organizations may have less time to assess and patch weaknesses before they are found or reused by others. OpenAI’s Linux figures, if independently borne out, suggest that a single automated effort can generate a substantial number of working exploits. They do not show that those exploits were used against real systems.

The benefits and risks are not limited to model providers. Security teams may use similar tools to audit code and verify fixes, while attackers could also apply increasingly capable systems to vulnerable software. A separate Epoch AI result cited in the source material says a freely downloadable model generated working attacks from published Firefox fixes for about $1,000 per Firefox version. That is a distinct reported experiment, not evidence that GPT-6 Cyber’s Linux results have been replicated.

For Anthropic, METR’s analysis raises a different question: whether AI tools are speeding up the research and engineering that produce later AI systems. If that interpretation is right, capability development could become harder to forecast using past rates alone. But the report’s restricted evidence means readers cannot independently examine much of the basis for its estimate.

Products Worth Considering

Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade

AmazonView Latest Price

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black

AmazonView Latest Price

Samsung Galaxy Tab A7 Lite 8.7 32GB Wi‑Fi Android Tablet – Long‑Lasting Battery, Compact Slim Design, Sturdy Metal Frame (US Version)

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

Two Reports, Different Evidence

OpenAI described its Linux campaign on November 4, before identifying GPT-6 Cyber as the model behind it the next day. The company says the test took place in mid-October 2026. The source material also compares it with an earlier OpenAI Linux run in July, which reportedly produced 166 privilege-escalation bugs for $513,690, or about $3,100 per exploit. Those figures come from the company’s reported runs, so the comparison is not an independent benchmark of performance across equivalent tests.

OpenAI’s account also cites Firefox security work. Mozilla’s public advisories reportedly credit OpenAI with 50 high-impact Firefox fixes in October, while OpenAI claims its work identified 81 previously unknown high-severity Firefox flaws. The advisories do not identify which model found each issue, so they cannot establish that GPT-6 Cyber discovered all the reported flaws. OpenAI rates the model “Critical” under its own Preparedness Framework, the framework’s highest cybersecurity tier; this is the company’s classification, not an outside certification.

METR’s report addresses a separate question from OpenAI’s security tests. It evaluates the pace of capability improvement at Anthropic, rather than measuring GPT-6 Cyber’s vulnerability-finding performance. Its estimate that AI accounts for much of the acceleration should be read alongside its disclosure that much of the underlying evidence is not available to the public.

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

— METR, in its report on Anthropic

Independent Verification Remains Limited

The available source material does not establish that an independent party reproduced OpenAI’s 948-exploit result, verified the count of 588 previously unknown flaws or audited the stated API costs. It also does not include a public response from Linux kernel maintainers. The figures should therefore be treated as OpenAI’s account of its own testing, not as independently confirmed totals.

The relationship between the Firefox figures is also unresolved: Mozilla’s advisories credit OpenAI with 50 high-impact fixes, while OpenAI reports 81 previously unknown high-severity flaws. The cited advisories do not map individual bugs to GPT-6 Cyber, and the source does not explain whether the two totals use the same definitions or counting method.

For METR’s findings, the main unresolved issue is transparency. The evaluator says Anthropic redacted a majority of the evidence it found most informative. The report offers METR’s conclusion and estimate, but the public cannot fully inspect the underlying material or independently confirm the proposed contribution of AI assistance.

Access and Evidence to Watch

OpenAI says GPT-6 Cyber is limited to approved Daybreak Red organizations, with access expected to extend to eligible customers over the coming weeks, according to reports cited in the source material. Further public detail about the model’s safeguards, additional testing and the handling of Linux findings could help clarify how the company’s claims are assessed. No schedule for an independent audit is given.

Readers should also watch for public statements from Linux maintainers and for security advisories that identify affected code, fixes or the provenance of reported findings. For organizations, the immediate practical step is to keep software updated and review security patches promptly; the reports do not establish that any particular Linux system has been attacked using the described exploits.

METR’s conclusions may be easier to assess if more underlying evidence becomes available or if later evaluations reproduce the reported change in Anthropic’s pace. Until then, the 1.9-times figure and the estimate that AI accounts for about 80% of acceleration remain METR’s analysis, with substantial evidence unavailable for public review.

Key Questions

What is GPT-6 Cyber?

OpenAI describes it as a cybersecurity-focused version of GPT-6 Astra for authorized security work. Access is restricted to organizations approved for its Daybreak Red programme.

Did GPT-6 Cyber produce 948 Linux exploits?

OpenAI says the model powered a mid-October test that produced 948 working local privilege-escalation exploits, 588 of which it says were previously unknown. The figures have not been independently verified in the cited material.

What did METR find about Anthropic?

METR reported that its main measure of Anthropic’s capability growth increased about 1.9 times faster after October 2025 than before. It estimates AI may account for about 80% of the acceleration, but says much of its most informative evidence was withheld.

Does the reporting show that Linux systems were attacked?

No. The reported exploits came from testing, and the source material does not say they were used in real-world attacks. It also reports no public comment from Linux maintainers.

What should software users do now?

Install security updates promptly and follow notices from software vendors and system maintainers. The reports do not identify a specific active attack requiring users to take a different action.

Source: ThorstenMeyerAI.com