GitHub App — Free for open source

AI agents are filing issues in your repo.

Automated agents now submit issues, open pull requests, and post comments at scale across GitHub. Assay detects them before they waste your maintainers' time — no configuration, no code changes.

One click. No account required. Works on any public repository.

13 Detection signals
3 Signal tiers by false-positive risk
0 False positives from Tier 1 signals
<1s Time to score each submission

Every submission scored before your team sees it.

When a new issue, PR, or comment arrives, Assay runs it through 13 behavioral signals — linguistic patterns, temporal cadence, identity fingerprinting.

Real contributors pass silently. Agents get flagged, labeled, and documented with a full signal breakdown so you can make the call.

Clean submissions are never touched. Assay only speaks when it has something to say.

A
assay bot commented just now
Assay: Likely AI-generated
84/100
Confidence: high

Signals detected

  • Sentence uniformity: CV=0.18 (human avg >0.40)
  • AI SDR opener formula: research hook + pivot + hard CTA
  • LLM vocabulary: 4 characteristic phrases detected
  • Heartbeat cadence: 3 prior submissions at 60-min intervals
ai-generated

Assay posts this comment automatically. Your team decides what to do next.

Three steps. No configuration.

01

Install the GitHub App

One click. Select which repositories to protect. No code changes, no YAML, no secrets to manage.

02

Assay scores every submission

New issues, PRs, and comments are scored against 13 signals — linguistic patterns, temporal cadence, identity analysis. Results in under a second.

03

Findings posted automatically

Suspicious submissions get a comment with the full signal breakdown and an ai-generated label. Clean submissions pass silently.

Three layers of detection.

Signals are organized by false-positive risk. High-confidence signals flag immediately. Lower-confidence signals require corroboration before Assay speaks. The result: almost no noise, and clear signal when it matters.

Tier 1 — Zero false positive
Dead giveaways
A single match here is enough. These patterns cannot appear in genuine human writing. Any one of them fires the highest confidence score.
  • Prompt leakage — unfilled template tokens like {{first_name}}
  • Agent metadata artifacts in submission headers
  • Honeypot challenge-response match
Tier 2 — Low false positive
Behavioral patterns
Humans are irregular. Agents are not. Two or more of these signals firing together is strong evidence of automated submission.
  • Heartbeat cadence — submissions at machine-consistent intervals
  • Send-time anomaly — unnaturally optimized timing
  • Superhuman response speed — replies within seconds
  • Ghost author — algorithmically generated username patterns
Tier 3 — Score boosters
Linguistic fingerprints
LLMs write in recognizable patterns. These signals boost the score when Tier 1 or Tier 2 signals are already present — never sufficient alone.
  • LLM vocabulary — characteristic phrase frequency
  • Sentence uniformity — unnaturally consistent length distribution
  • AI opener formula — research hook + pivot + CTA structure
  • No human artifacts — absence of typos, contractions, informality
  • Reply mirroring — addresses every point in submission order

Designed to protect real contributors.

Tier 3 signals cannot flag a submission on their own. Perfect AI vocabulary and sentence uniformity without any behavioral signals will score in the 40–60 range — marked "uncertain," not "agent." A non-technical human contributor who writes carefully will pass silently. The tiered architecture makes false positives rare by design.

Free for open source. No strings.

Install Assay on any public repository in one click. No account required, no configuration, no API keys, no YAML. Assay starts scoring every new submission immediately.

Free for public repositories
Issues, PRs, and comments
No data stored after scoring
Uninstall in one click