OpenAI’s 4 November 2026 report describes a mid-October Linux test using an upcoming cyber model: 2,308 candidates, 1,724 reproduced vulnerabilities, 1,169 working local privilege-escalation exploits, and 948 unique vulnerabilities counted by distinct patches after variant analysis.
API credits cost $1,213,430, not necessarily the full project budget. Humans reviewed outputs; hundreds of reports and patches were submitted, with more being coordinated.
These are reported results, not our independent replication. Generated patches are not proof of deployed fixes. Check your distributor’s advisories.