Environments

On request, we send AI labs and research teams the environments, each a packaged image with its grader and trigger, and run transcripts.

Unpatched

Vulnerabilities XOR researchers discovered and disclosed in a confidential inference stack.

Attack · Unpatched

Any trigger: the exploit reached the diagnostic. Full exploit: the grader accepted it. Both over every graded run.

Full exploit 0.0%Any trigger 63.5%details unpublished, the defect is not disclosed · across systems' graded runs · newest run 10 September 2026

Defend · Unpatched

Any credit: partial or full credit. Full fix: full credit. Both over every graded run.

Full fix 11.8%Any credit 21.2%details unpublished, the defect is not disclosed · across systems' graded runs · newest run 10 September 2026

Planted vulnerabilities in cryptography libraries with no public fix: a patch must stop every attack and keep users in.

Patch that keeps users in · the defender role · Cryptography

Full fix 68.2%Any credit 81.8%across systems' graded runs · newest run 11 September 2026

Attack · Cryptography

Full exploit 0.0%Any trigger 50.0%across both environments; see split · newest run 11 September 2026

Memory-safety vulnerabilities in low-level software with public fixes: systems patch them and exploit them.

Patch · Vulnerability repair

Each system repairs the vulnerability; the grader replays the trigger and rebuilds.

Pass rate 74.4%across systems' graded runs · newest run 30 April 2026

Exploit · Vulnerability repair

The same vulnerabilities, graded on whether the system's exploit works.

Pass rate 84.4%across systems' graded runs · newest run 30 April 2026