How A $1,279 Exploit Fits Into The GPT-6 Cyber And METR Findings

🔍 Read the full analysis: How A $1,279 Exploit Fits Into The GPT-6 Cyber And METR Findings on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

OpenAI said on November 5 that its newly named GPT-6 Cyber model powered a mid-October Linux kernel test that produced 948 working privilege-escalation exploits, at an API cost of about $1,279 per exploit. In a separate report published the same day, evaluator METR said Anthropic’s measured model progress had accelerated, with much of the increase likely driven by AI assistance to its own researchers. Both developments raise questions about how quickly security flaws can be found—and how much of the evidence is independently checkable.

OpenAI’s figures describe an automated campaign against part of the Linux kernel, software used across servers, computers and other devices. The company said the agents scanned about 2.8 million lines of code, or roughly 7% of the project, and examined 2,308 candidate findings. OpenAI reported reproducing 1,724 vulnerabilities in test machines before producing 948 working local privilege-escalation exploits. It said 588 exploits involved previously unknown flaws. These figures come from OpenAI’s account; the supplied material does not include an independent audit of the full run.

The reported API bill was $1,213,430, which works out to about $1,279 for each of the 948 exploits. A previous OpenAI run in July reportedly produced 166 Linux privilege-escalation bugs for $513,690, or about $3,100 each. Those totals suggest a lower reported cost per result in the later campaign, but they do not establish that the two runs used identical methods, targets or criteria. OpenAI named GPT-6 Cyber as the model behind the October run when it announced the model on November 5.

OpenAI describes GPT-6 Cyber as a cyber-specialized version of GPT-6 Astra, built for authorized security work such as vulnerability discovery and exploit testing. Access is restricted to organisations it approves for the Daybreak Red programme, which covers offensive security testing. The system card describes requirements including multifactor authentication, managed devices and restricted API access. OpenAI assigns the model its highest internal cybersecurity preparedness rating, “Critical”; that is the company’s assessment, not an independent certification.

Falling Costs Put Pressure on Patch Teams

The reported cost matters because it suggests that AI-assisted vulnerability research may produce more results for a given amount of API spending than earlier runs. The figures are not a full measure of the cost of finding, validating, reporting and fixing a flaw, and they do not show how often the same approach works across other software. Still, a lower reported cost per working exploit could make large-scale security testing more accessible to both defenders and attackers.

That puts pressure on organisations to find and fix flaws quickly. If software weaknesses can be identified and turned into working exploits faster, the period between discovery and abuse may narrow. The source also describes an Epoch AI experiment in which a freely downloadable model, GLM 5.3, reportedly turned published Firefox fixes into working attacks for about $1,000 per Firefox version. That is a separate result, not evidence that GPT-6 Cyber did the same work or that every organisation can reproduce it.

METR’s report adds a second concern: AI may be helping labs improve the systems that can then assist with security research. METR estimated that Anthropic’s main measure of model capability grew 1.9 times faster after October 2025 than before, and attributed about 80% of the acceleration to AI helping Anthropic researchers. The report’s estimate is not a direct measurement of every research task or a guarantee of future progress. Taken together, the findings indicate potential pressure on both software defenders and oversight of AI development.

Products Worth Considering

Therabody Theragun Mini 3rd Gen Massage Gun (Travel Essential, Black)

AmazonView Latest Price

Philips Sonicare 5950 Rechargeable Electric Toothbrush with 2 Brush Heads, Pink

AmazonView Latest Price

Oura Ring 5 Smart Ring – Stealth, Size 11

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

Two Reports, Different Evidence

The Linux results and METR’s analysis are related by the broader question of AI-enabled progress, but they are separate findings with different evidence. OpenAI’s exploit counts and costs come from its own description of a controlled campaign. The supplied account says Mozilla’s October security advisories credit OpenAI with 50 high-impact Firefox fixes, but do not identify which model found each issue. It also says no Linux maintainer had commented publicly at the time of the report. Those details offer some external context, but do not independently confirm the Linux totals.

GPT-6 Cyber is not generally available, according to OpenAI’s stated access policy. It is limited to vetted organisations approved for Daybreak Red, with access through the API and Codex. OpenAI says it monitors usage and can suspend or revoke access. The company’s “Critical” designation reflects its own Preparedness Framework and should not be read as an external finding about the model’s performance or risk.

METR’s report concerns the pace of improvement at Anthropic, rather than the Linux exploit campaign. According to the supplied material, METR spent eleven weeks examining data from Anthropic. It reported a faster rate of improvement on its main measure after October 2025 and estimated that AI tools account for much of the acceleration. METR also said Anthropic redacted a majority of the evidence it considered most informative, limiting what readers can review themselves.

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

— METR, as quoted in the supplied report

Independent Checks Remain Limited

The central uncertainty is how much of OpenAI’s reported Linux result can be verified outside the company. The supplied material gives OpenAI’s counts for scanned code, candidate findings, reproduced vulnerabilities and working exploits, but does not provide an independent review of the complete dataset or methodology. It is also not clear from the available information which individual flaws were new, how they were disclosed to maintainers, or how many have been patched.

The comparison with July also has limits: the source gives the two runs’ reported costs and exploit counts, but does not establish that they used the same procedures or success criteria. Likewise, Mozilla’s reported Firefox acknowledgments do not establish which model found particular vulnerabilities. OpenAI’s access controls are described by the company; the supplied material does not provide an outside assessment of how effectively they prevent misuse.

METR’s conclusions also carry a stated evidence limitation. The evaluator said Anthropic redacted a majority of the evidence it found most informative. Its estimate that AI accounted for about 80% of the acceleration is therefore an assessment based on information METR reviewed, not a publicly reproducible calculation from a complete data record. The source material does not provide a public response from Anthropic addressing that estimate.

Disclosure and Patching Will Matter

The immediate next step for the Linux findings is disclosure and remediation. The supplied source says no Linux maintainer had publicly commented when it was written, and it does not establish a timetable for fixes or identify which reported flaws have already been addressed. Readers should watch for maintainer statements, security advisories and patch releases that can clarify the scope and status of the vulnerabilities.

OpenAI says GPT-6 Cyber access is restricted to approved Daybreak Red organisations, with wider access for eligible customers expected over the coming weeks, according to reports cited in the source material. Further details on access, monitoring and the model’s use may clarify how the company intends to limit testing to authorized work. Independent technical review of the Linux results would also help establish how the reported exploit count should be interpreted.

For organisations and individual users, the practical response remains straightforward: apply software security updates promptly and follow vendor and maintainer guidance. The reports do not show that a particular person or organisation is under attack, but they reinforce the need for timely patching as automated security research develops. METR’s findings may be clearer if more supporting evidence becomes available or if later evaluations test whether the reported pace of improvement continues.

Key Questions

What was the $1,279 figure?

It is the approximate API cost per working Linux privilege-escalation exploit, calculated from OpenAI’s reported $1,213,430 bill divided by 948 exploits. It does not include every possible cost of research, validation or fixing vulnerabilities.

Did an independent body verify the 948 exploits?

The supplied material does not identify an independent audit of OpenAI’s full Linux campaign or its totals. OpenAI reported the figures, while the source notes that no Linux maintainer had commented publicly at the time.

Who can use GPT-6 Cyber?

OpenAI says access is limited to organisations it approves for its Daybreak Red programme, which is intended for offensive security testing. The company describes identity, device and access-control requirements.

What did METR report about Anthropic?

METR said Anthropic’s main measured rate of model improvement was 1.9 times faster after October 2025 than before. It estimated that AI assistance accounted for about 80% of the acceleration, while noting that much of its most informative evidence was redacted.

What should software users do now?

Install security updates promptly and follow guidance from software vendors and maintainers. The reports do not establish that a specific user is being targeted, but they underline the value of reducing the time systems remain unpatched.

Source: ThorstenMeyerAI.com