$1.2 million for 948 exploits: the security bill behind AI scans

Monday begins with an open security alert. Updates are sitting in your inbox, an external provider is waiting for approval, and no one is quite sure which systems to check first. OpenAI says GPT-6 Cyber generated 948 working exploits for local privilege escalation in a Linux study. The run used $1,213,430 in API credits, according to the company. The figure looks like a price list. It is not. What it actually tells you depends on what your team can check and the limits you set for the model.
The figure is no price list
OpenAI gives the cost of one run as $1,213,430 in API credits, with 948 local privilege escalation exploits generated from its Linux research. That works out at about $1,279 per exploit. But this ratio describes one specific research process. It is not a standard cost per vulnerability found.
The run analysed 2,308 possible Linux kernel findings and dynamically reproduced 1,724 vulnerabilities. Here, “dynamically reproduced” means checking whether a suspected issue can actually be replicated in a live, controlled environment. The pipeline generated exploits only after that step. OpenAI says costs depend on the models, target configuration and workflow. Additional technical orchestration could bring them down, the company says.
For a small or medium-sized business with a lean IT team, the more useful question is not whether $1,279 per finding would be affordable. It is this: who can sift genuine risks from a long list of possible findings, and who responds if one of them suddenly calls for an urgent patch?
From suspicion to verified finding
An AI model scanning source code for vulnerabilities starts by producing suspected issues. A finding is not proof that someone can exploit it. OpenAI's described pipeline therefore works in stages: it selects areas of code, looks for possible flaws, checks them in isolation and then creates what is known as a proof of concept. That demonstrates that a flaw can be reproduced under specific conditions.
A human security researcher reviews the results at the final stage. The pipeline can also suggest patches, meaning code changes intended to fix a flaw. A suggestion saves time only once it has been checked by an expert, tested and approved for your environment. A patch applied in haste could cause new problems or conflict with local changes.
OpenAI says it shared 304 critical Linux findings, along with possible patches, with maintainers. That shows how important the handover from finding to fixing can be. For your business, it means the AI run is not the end of the security process. It can create a new queue of work that someone has to own.
Testing attacks is getting cheaper too
When a tool can find vulnerabilities and generate exploits, it changes the security question on both sides. A local privilege escalation exploit could let a user with limited initial permissions gain root access to a Linux system. That is a high level of privilege. For businesses with internal servers, development environments or shared systems, this kind of flaw is more than a technical footnote.
OpenAI also says GPT-6 Cyber identified 81 previously unknown, high-severity vulnerabilities in Firefox. Mozilla has already released fixes for many of them, according to the company. In addition, 30 known Firefox issues were rated as more serious than previously thought. These are results OpenAI reports from its research, not a guarantee that every model will test every browser or business system with similar success.
In ExploitGym, a test environment for tasks involving vulnerability exploitation, GPT-6 Cyber scored 48.5 per cent, according to OpenAI. Its predecessor, GPT-5.6 Cyber, scored 32.1 per cent. That progress matters, but the score does not tell you how secure your specific application is. A pilot needs to close that gap between a test result and your own environment.
Controls are part of the tool
A model should not search everywhere during a security test simply because it has the technical ability to do so. OpenAI says GPT-6 Cyber did not attack services outside the defined scope in 99.93 per cent of its tests without any additional instruction. With a purpose-written deployment instruction, the figure was 100 per cent.
Those are the company's test results, not a guarantee for every configuration. A test scope set incorrectly could include systems belonging to someone else. Even a well-intentioned check can disrupt live services. Written approval, clearly defined targets and an isolated test environment need to be in place before the first run, not just discussed afterwards.
OpenAI rates GPT-6 Cyber as “Critical” for cybersecurity in its own Preparedness Framework, one level above the “High” rating for GPT-5.6 Cyber. That is the company's assessment within its own security framework. It is not an independent certification and no reason to give the model a free hand. The more powerful the search, the more precisely you need to define its scope.
What it means for your business
A business with its own security team can fit this approach into existing testing and approval processes. In a small or medium-sized business, the work often falls to a handful of people who are already juggling updates, supplier queries and day-to-day operations. The extra checking is no footnote. It determines whether an AI finding speeds up remediation or creates yet another list in someone's inbox.
Start with a small, approved test rather than your entire IT estate. Choose a system whose owner has authorised the test, and record which targets the model is allowed to examine. Decide in advance who will verify a finding, who will test a patch and who will approve it.
During the pilot, track three things: the time and cost of the model run, the number of confirmed findings, and the effort needed to check and fix them. Agree a stop condition in case the model leaves the test scope or disrupts operations. By next Monday, it should be clear who will assess the first finding and which test system will receive a patch. Only then can you judge whether the process actually takes work off your team's hands.
- Approve the targets and test window in writing.
- Keep the test environment separate from production.
- Name someone to review findings and approve patches before the test begins.
- Compare the total cost and time spent with your current testing process.
The patch is not the end
The value of a vulnerability finding also depends on how quickly maintainers and operators can respond. OpenAI says that while its Linux scan was under way, more than 100 local privilege escalation vulnerabilities were patched or had patches in preparation within ten days. About 300 more were being discussed publicly.
These are observations from OpenAI's research, not a promise that your provider or software supplier will act just as quickly. Still, expectations are changing: if more testing produces more concrete findings, clear ownership and a workable patch process matter more. The bottleneck could shift from finding issues to assessing, testing and applying fixes.
So ask your IT team or security provider more than whether AI can find vulnerabilities. Ask how a confirmed finding will be recorded, who will prioritise it and how you will know when a patch is available for your version. Without those answers, even a fast discovery is still just an open ticket.
Frequently asked questions
What is GPT-6 Cyber?
GPT-6 Cyber is a model introduced by OpenAI for cybersecurity research tasks. OpenAI says it can investigate software vulnerabilities, generate exploits as proof of concept and help suggest patches. Results from the published tests do not automatically show how well the model will work in a particular business environment.
How much does an AI-powered Linux security scan cost?
OpenAI puts the cost of one of its Linux research runs at $1,213,430 in API credits. The company says it generated 948 local privilege escalation exploits. These costs apply to the models, target configurations and workflows used in that research. OpenAI says additional technical orchestration could reduce the cost, so this is not a standard price for an AI security scan.
Can AI find security vulnerabilities reliably?
AI can identify possible vulnerabilities and test whether they can be exploited in a controlled environment. In OpenAI's Linux research, the company says 1,724 of 2,308 possible findings were dynamically reproduced. That is a specific research result, not a guarantee of reliability on other systems. Suspected issues, exploits and patch suggestions should be reviewed by specialists.
What controls does a business need for AI security testing?
Before the test, set out in writing which systems may be checked and when the run will take place. Use an isolated test environment and decide who will verify findings, review patches and approve changes. Also agree to stop the test if it leaves its defined scope or disrupts live operations.
Can GPT-6 Cyber fix the vulnerabilities it finds?
OpenAI says its Linux pipeline can generate candidate patches. A suggestion is not an approved fix. Someone needs to check that the patch resolves the flaw and will not cause new problems in your environment. It then needs a controlled test and clear approval before it goes into production.
Interested?
Want to know what this means for your company?
Sources
- GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents · openai.com, 2026-11-05
- A thousand fewer paths to root · openai.com, 2026-11-04
- GPT-6 Cyber, OpenAI automatizza la caccia ai bug: cosa cambia per le imprese - AI4Business · ai4business.it, 2026-11-06
- En medio de una crisis de ciberseguridad, OpenAI estrena GPT-6 Cyber, su nuevo modelo "crítico" · xataka.com, 2026-11-06
This article was researched and written with AI assistance and automatically checked against the sources listed before publication. More on our AI transparency
Free guide · PDF
The AI Compass for small and mid-sized businesses
9 jobs your team can hand over by Monday.
Get it free
Klarframe AI Readiness Check
How ready is your company for AI?
About ten minutes, instant result with your stage, a 90-day plan, suitable solutions and notes on the EU AI Act.


