agent-dev-workflow

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). In references/bin/gc, the runtime ingests outsider-authored free text by running git -C "$rwt" cherry "$base_ref" HEAD and (optionally) gh pr list --head "$br" --state merged --json ... --jq ..., where the LLM-input comes from command output that can include untrusted branch/PR metadata authored externally.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The scripts accept gs:// snapshot URLs (e.g. "gs://bucket/x.sql" / "gs://${BUCKET}/latest.sql") and at runtime run commands like gcloud storage cat "$snapshot" | app_strip_snapshot | app_psql -d ..., which fetches remote SQL and executes it in the database.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 1.00). The skill instructs adding and running project scripts that stop/start Docker containers, delete derived databases, remove worktrees/branches, and kill running sessions (via bin/gc, bin/stop, bin/cleanup), which are destructive, machine-scoped state changes that can modify or remove data and processes.