I attacked my own repo — my PR bot blocked the attack before merge

Description and use

AgentGuard detected and blocked 12 hidden attacks in agent configuration, prevented a dangerous push via GitHub Secret Scanning, and stopped an attack in a live PR without human intervention.

Problem

Coding agents can execute malicious instructions hidden in repo files, risking security breaches before code review.

Outcome

AgentGuard successfully detected all crafted attacks, prevented a malicious push via GitHub secret scanning, and blocked an attack in a live PR, demonstrating effective security gating for coding agents.

Workflow

• 1. Build AgentGuard CI gate to scan AGENTS.md, skills, MCP servers, hooks.
• 2. Inject 12 targeted attacks into a test repo.
• 3. Run AgentGuard unit tests to confirm detection of all attacks.
• 4. Scan 30 public repos for AGENTS.md; identify 5 with critical findings.
• 5. Manually verify findings and refine rules to eliminate false positives.
• 6. Push a test commit containing a fake Slack token; GitHub secret scanning blocks the push.
• 7. Deploy AgentGuard GitHub App (Pro) to comment on PRs and run check‑runs.
• 8. Create a harmless PR; observe green check‑run and merge allowed.
• 9. Create an attack PR; observe failure check‑run and merge blocked.

Evidence

AgentGuard detects all 12 — deterministically, no LLM oracle — with 36 unit tests that prove it forever (12/12 detected, 0 false positives on the clean control repo).

Source

GitHub's secret scanning caught it and blocked the push.

Source

PR #2 (attack: hidden instruction with exfil in AGENTS.md): comment Note E (1 findings), check-run failure → merge blocked.

Source