[detection-stats] Detection stats for github/gh-aw - 2026-08-22

[github-actions]

962 external-detector runs analysed on 2026-08-22; detection-job error rate 0%; soft failures 71; threat rate 0.47% (2/429 runs with a verdict).

Summary

gh-aw detection statistics - 2026-08-22 (UTC)

Repository: github/gh-aw
Window: 2026-08-22T00:00:00Z .. 2026-08-22T23:59:59Z
API requests: 2938, rate-limit pauses: 1
Data complete: yes

Totals

Metric	Count
Workflow runs in window	3781
Agentic runs (*.lock.yml)	1955
Runs with a detection job	980
... using the external detector	962
... using the built-in detector	18
... detector could not be determined	0
Agentic runs without a detection job	975

All rates below are over the external detector population (962 runs). Since gh-aw #54111 the external detector is the compile-time default; a run counts as external when its detection job showed the Install threat-detect binary step, when another run of the same workflow did that day, or when the workflow is .lock.yml and no run showed the built-in shape (a completed detection job with steps but no marker). Runs on a workflow that opted out with features: gh-aw-detection: false count as built-in.

Detection job outcomes

Outcome	Count	Rate
success	503	52.29%
skipped	459	47.71%

Error rate (failure/timed_out/action_required): 0%

Verdict availability

State	Meaning	Count
present	detection artifact downloaded and parsed	429
absent	detection job ran but published no artifact (soft failure)	71
skipped	detection job was skipped or was still running (nothing to fetch)	459
unreadable	artifact zip could not be unpacked	3

Green detection jobs that published no verdict: 71 (detection steps are continue-on-error, so a missing verdict artifact is the only reliable signal for these).

Detection results

Result	Count
Runs with a parsed verdict	429
Clean (no threat)	427
Any threat	2
prompt_injection	1
secret_leak	0
malicious_patch	1

Threat rate (of runs with a verdict): 0.47%

Reasons reported by gh-aw

From the [aw] Detection Runs tracking issue (warning/failure conclusions only), restricted to runs in the external-detector population above.

Reason	Count
agent_failure	73
threat_detected	2

By workflow

Workflow	Runs	Failed	Cancelled	Skipped	No verdict	Threats
Q	151	0	0	151	0	0
PR Sous Chef	78	0	0	0	11	0
Deployment Incident Monitor	61	0	0	61	0	0
Issue Monster	46	0	0	21	6	0
Test Quality Sentinel	43	0	0	0	2	0
Design Decision Gate 🏗️	42	0	0	1	3	0
Impeccable Skills Reviewer	42	0	0	0	2	0
Matt Pocock Skills Reviewer	42	0	0	0	2	0
PR Code Quality Reviewer	42	0	0	0	2	0
Ponytail Reviewer	42	0	0	0	4	0
Avenger	22	0	0	16	2	0
Daily Go Test Parallelizer	12	0	0	9	0	0
PR Description Updater	12	0	0	4	1	1
Squad — @copilot resolve the merge conflicts on this branch.	10	0	0	10	0	0
Auto-Triage Issues	7	0	0	1	1	0
Contribution Check	5	0	0	1	0	0
Code Scanning Fixer	4	0	0	0	1	0
Workflow Generator	4	0	0	4	0	0
[aw] Failure Investigator (6h)	4	0	0	0	1	0
Deep Report	3	0	0	0	0	0
PR Triage Agent	3	0	0	0	0	0

260 further workflows omitted; see stats.json.

Notable runs

Selected examples (79 notable runs total in the full data; showing a subset here, see stats.json for the complete list):

Workflow	Run	Job conclusion	Verdict	Threats	Reported reason
Daily Community Attribution Updater	32546873672	success	present	malicious_patch	threat_detected
PR Description Updater	32548155220	success	present	prompt_injection	threat_detected
Agent Job Health Monitor	32604582938	success	absent	-	agent_failure
Agentic Workflow Audit Agent	32598934886	success	absent	-	agent_failure
Auto-Triage Issues	32572717304	success	absent	-	agent_failure
Avenger	32597081051	success	absent	-	agent_failure
Avenger	32599906451	success	absent	-	agent_failure

(remaining rows omitted for brevity — see stats.json for the full list)

Change since 2026-08-21

Metric	2026-08-21	2026-08-22	Delta
External-detector runs	1001	962	-39
Error rate	0.3%	0%	-0.3 pp
Soft failures	54	71	+17
Runs with verdict	481	429	-52
Any threat	2	2	0
Threat rate	0.42%	0.47%	+0.05 pp

(Fewer than 7 history entries exist, so no 7-day mean column yet.)

Watch list

• PR Sous Chef — 0 failed, 11 without a verdict, out of 78 runs
• Issue Monster — 0 failed, 6 without a verdict, out of 46 runs
• Ponytail Reviewer — 0 failed, 4 without a verdict, out of 42 runs
• Design Decision Gate 🏗️ — 0 failed, 3 without a verdict, out of 42 runs
• Test Quality Sentinel — 0 failed, 2 without a verdict, out of 43 runs

Collected by: https://github.com/github/gh-aw-threat-detection/actions/runs/32617093541
Full data: the detection-stats-32617093541 artifact on that run.

Generated by Detection Stats Daily · auto · 35.6 AIC · ⌖ 9.92 AIC · ⊞ 11.3K · ◷