Independent benchmark · live data
Scanner Recall Explorer
Four reads on the same scanners. Detection on a home-ground slice versus a held-out one; which inert capabilities each tool flags; how we replicate the one other public run on the same set; and how a single Cisco binary reads anywhere from 2.5% to 92.7% across benchmarks.
Presented with one inert capability and no payload, which scanners flag it? Every sample is a harmless canary that does exactly one thing — the question is recognition, not malice.
The one other public static run on this exact held-out set is a Cisco-side study (Vineeth Sain's HF space). At a matched severity bar our numbers land within a few points of theirs; the only gap is at each tool's shipped gate, where the two count different things.
Computed live from leaderboard.json. Vendor static
scanners are the neutral comparison; skill-lint is our own linter (muted,
not ranked) and the Tencent rows are LLM-judged. External numbers are cited
to the paper or page they come from. Regenerate ours with python -m score.report.