HoF-Bench: leaderboard

Metric: CVEs rediscovered at pass@4 (out of 95), with the generated-context stage, on 95 public AI-discovered CVEs in eight repositories pinned at vulnerable commits, found by a minimal fixed LLM scanner scaffold with four repeated passes and multi-round triage, credited only when a detector-blinded GPT-5.6 terra judge matches code path, root cause, attack condition and impact; medium reasoning effort where supported. Source: arxiv.org. Saturation forecast: Around May 2027. 10 models tracked.

Top models

#ModelScore
1GPT-5.6 Luna (Medium)65
2DeepSeek V4 Flash62
3GPT-5.4 Mini (Medium)58
4Gemini 3.5 Flash (Medium)56
5GPT-OSS-120B52
6Qwen 3.6 35B A3B50
7Gemini 2.5 Flash (Medium)47
8GPT-5.4 Nano (Medium)46
9GPT-OSS-20B40
10Mistral Small 438

Interactive version: theaggregate.ai/benchmark?slug=hof-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.