OpenAI's AI Finds 948 Linux Kernel Root Exploits at $1,279 Each

On November 4, OpenAI published results of a scan testing whether regular user accounts could breach the Linux kernel to gain root privileges.

October 2024 Scan Results

The October scan used an updated model and produced 948 results. The total API cost was $1,213,430 , yielding an average cost of $1,279 per result (1,213,430 ÷ 948). This represents a dramatic increase over earlier runs: the June scan using GPT-5.5 Cyber produced only 24 results, while the late July scan using GPT-5.6 Sol produced 166 results at a cost of $513,690 . The October run delivered more than five times the July count.

Candidate Verification and Accuracy

Across the October run, 2,308 candidates were generated. Of these, 1,724 were reproducible ( 74.7% ). Analysis errors — where the model treated an impossible condition as true — accounted for only 3.1% . These figures come from OpenAI's own execution logs, not from independent replication.

Early Development and Process Improvements

During early development, OpenAI reported 200 successes. Human review revealed that only 20 actually achieved root via the specified vulnerability; the rest exploited other issues. In response, OpenAI tightened the task definition and added an independent verification step to confirm that the designated vulnerability was used.

Remediation Timeline and Upstream Status

The scan and human analysis took approximately 10 days . By the time of reporting, upstream maintainers had already merged 61 fixes, 43 had pending patches, and 299 were under public discussion. OpenAI states they submitted hundreds of vulnerability reports and patches.

kernelCTF Benchmark Run

A separate run on the kernelCTF platform targeted a hardened configuration (not a standard desktop). It cost $32,936 , ran for 112 hours , and produced 25 information-leak proofs and 21 root-privilege exploits.

Cumulative 2026 Results

As of the article's writing in 2026, the project had accepted 26 confirmed root-privilege vulnerabilities. OpenAI's own runs accounted for 21 of these. The two counts are not directly additive and should not be treated as a single tally.

Interpretation Caution

The 948 figure should not be read as 948 distinct unpatched holes on deployed systems. Rather, a single scan run can yield that many results, some of which are already in public discussion or have patches submitted. The number estimates the yield of one scan, not the length of a backlog.

Source: https://openai.com/index/a-thousand-fewer-paths-to-root/

Code example

👀点击关注
AI前沿快速Follow