OpenAI describes a Linux research pipeline that produced 2,308 candidate findings, dynamically reproduced 1,724 and developed 948 local privilege-escalation exploits on its stock desktop target. It generated candidate patches for review, rather than reporting that all findings were novel or all fixes had shipped. Hardened-target results are a separate cohort.
Why it matters
It exposes the stages between a high-volume model scan and maintainers accepting useful security fixes.
Key facts
As stated in the sources, with where to find them.
- The mid-October run produced 2,308 candidates, 1,724 dynamically reproduced findings and 948 local privilege-escalation exploits on the stock desktop target. These stages should not be treated as interchangeable bug or CVE counts.Summary
- OpenAI reports about $1.2 million in API cost for the run. This is model usage, not a complete measure of human review and upstream remediation costs.Cost analysis
- The pipeline produces candidate patches for human researchers and Linux maintainers to review; successful exploit reproduction does not establish a shipped fix.Patch generation
- OpenAI acknowledges overlap with bugs patched, staged or publicly discussed by other contributors during the scan. The exploit total therefore does not establish 948 uniquely new vulnerabilities.Maintainers are moving quickly to patch vulnerabilities
- An adapted kernelCTF target yielded 21 local privilege-escalation and 25 information-leak proofs of concept. The test omitted SLAB_VIRTUAL because of kernel-version incompatibility, so it did not include every kernelCTF mitigation and is separate from the stock desktop run.What defenders can learn from exploit costs
Findings that cite this record
Key questions this bears on
- Is AI shifting the balance between finding and fixing vulnerabilities?Discovery is ahead. Verification, triage, coordinated disclosure, safe repair and getting fixes to users remain bottlenecks between AI findings and protection.
- How are attackers using AI agents in real operations?To run large parts of intrusions under human direction, including espionage, credential theft and extortion; reported autonomy and success vary.
Sources
Related records
Oct 21, 2026
Nov 4, 2026
Oct 28, 2026
Oct 13, 2026
Oct 12, 2026
Oct 5, 2026