Clawmont AI agent security benchmark.

Measured runtime-security results with their limits kept visible. The benchmark below separates performance on a fixed adversarial corpus from performance on fresh, unseen attacks. The scorecard that follows is the latest published release snapshot, not a guarantee of protection.

Measured outcomes

Fixed corpus and fresh deterministic test, side by side

fixed 2026-06-10 / R4 2026-06-17

False positives: 0; Across 62 clean negatives
Fixed-corpus detection: 78.7%; 1,829 of 2,324 adversarial vectors
Novel deterministic detection: 33.11%; 1,033 of 3,120 fresh R4 attacks
Unexpected fixed-corpus bypasses: 0; 495 known misses remain in the rate

The 78.7% result comes from a fixed corpus the detectors were iterated against, so it overstates performance on attacks they have never seen. Fresh, unseen plain-English attacks measured 33.11% on the deterministic checks only. These are best-effort test outcomes, not a guarantee.

machine-readable JSON methodology and interpretation

Latest published release scorecard

This scorecard is a separate release snapshot. Its version and date are shown below and may differ from the benchmark runs above.

Release scorecard grade

—

—/10

as of — · v—

attack resistance—

tests passing—

modules covered—

Score over time

A/B (≥8.0) C (6.0–7.9) D/F (<6.0)

Hover a point to see what changed that day.

Category breakdown

Recent changes

Short, non-sensitive summaries only. Exploit details are held back by policy — see disclosure.