Skip to main content
XOR
UNPATCHED

Unpatched

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

The environments and their graders are withheld; the leaderboards show aggregate results.

MethodAdd your systemResearch note: Cyber frontier eval: GPT-6 Astra and Claude Fable 5.1 repair only a third of the flaws in realworld conditions

Attack

7 systems · newest run 10 September 2026

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

Ordered by Full exploit, highest first. Ties sort by Any trigger, then by name. Unranked. Attack leaderboard, newest run 10 September 2026.
Claude Code Opus 5 max95.0%0.0%
OpenCode Kimi K3 xhigh85.0%0.0%
OpenCode Qwen 3.8 Max xhigh84.2%0.0%
OpenCode GLM-5.3 xhigh78.9%0.0%
Codex GPT Daybreak Blue xhigh54.5%0.0%
Codex GPT-5.6 Sol xhigh40.0%0.0%
OpenCode Gemini 3.8 Flash xhigh25.0%0.0%
How we measured this
  • Any trigger. the exploit diagnostic reached its second rung: the planted flaw fired
  • Full exploit. the grader's exploit channel accepted the exploit in full
  • Cell. One run of one system on one environment. Every retained run counts as its own cell.
  • Order. Rows sort by any column; runs this close are not separated, and no rank is published.
  • Contamination. No public fix exists for these vulnerabilities; the grader tests whether the vulnerability still triggers.
  • Newest run. 10 September 2026 · Cite · Method

Defend

9 systems · newest run 10 September 2026

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

Ordered by Full fix, highest first. Ties sort by Any credit, then by name. Unranked. Defend leaderboard, newest run 10 September 2026.
Codex GPT-6 Astra xhigh41.2%35.3%
Claude Code Fable 5.1 high42.9%28.6%
Codex GPT-5.6 Sol xhigh29.2%16.7%
Claude Code Opus 5 max19.4%12.9%
Codex GPT Daybreak Blue xhigh14.8%11.1%
OpenCode GLM-5.3 xhigh20.0%8.0%
OpenCode Gemini 3.8 Flash xhigh27.6%6.9%
OpenCode Qwen 3.8 Max xhigh12.5%4.2%
OpenCode Kimi K3 xhigh7.1%3.6%
How we measured this
  • Any credit. the patch earned any credit from the grader, partial included
  • Full fix. the patch earned full credit from the grader
  • Cell. One run of one system on one environment. Every retained run counts as its own cell.
  • Order. Rows sort by any column; runs this close are not separated, and no rank is published.
  • Newest run. 10 September 2026 · Cite · Method
Unpatched | XOR