armstat — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited armstat (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Run from the project root against one or more turn-log NDJSON files:
node benchmark/bin/arm_stat.js runs/<run>/turns.A0.ndjson
node benchmark/bin/arm_stat.js runs/<run>/turns.*.ndjson # whole suite so farRead-only — it only reads files already on disk, never the running proxy, so it is safe to run while a benchmark arm is still driving.
turns — billable client turns (+ keep-alive treatment-ping records,counted separately and excluded from the billable tally), and any partial trailing line skipped on a live file.
httpStatus — status tally with a ✓ when all are 200, or a ⚠️ flag themoment any non-200 appears (early rate-limit / error detection mid-run).
ttlMode — 5m/1h tally (passthrough arms stay 5m).cache_read / cache_create — token sums, plus the per-turn first→lastcache_read (warm-cache build-up), and the ephemeral 5m-vs-1h split. The split is a direct knob check: the 1h-TTL arms (A2/A3/A5) should show ephemeral_1h > 0; A0/A1/A4 should not.
hit rate — token-weighted `cache_read / (cache_read + cache_create +input). A quick diagnostic only; the defensible headline is billable input reclaimed per turn with a CI, which the bench` analyzer computes.
thinkingBudget — distinct budgets seen and how many turns carried one(Haiku emits ~31999 regardless of --effort).
produced turns, hit no rate-limit (all 200), warmed its cache, and applied the knob it was supposed to (5m vs 1h).
bench: no CIs, no gap-bucket stratification, nopricing. Run bench for the committed study report.
Turn-logs carry per-turn token usage (not prompt content). Treat the run directory as sensitive; this tool prints only counts and token sums.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.