Attack Library

Browse seed attacks by category. These attacks form the initial population for campaigns.

About Seed Attacks

Seed attacks are the starting strategies that PentestLoop uses to begin testing your agent. Think of them as the "generation zero" of an evolutionary process. Each seed defines an objective (what to test), a persona (how to approach it), and opening messages (how to start the conversation).

During a campaign, these seeds are tested, scored, and the most effective ones are mutated and crossed over to create new, more sophisticated attack strategies. Over generations, the attacks evolve to find weaknesses that simple static tests would miss.

Categories

🛡️ Security

Attacks that test your agent’s ability to protect sensitive data, resist prompt injection, and maintain safety guardrails. These probe for data leakage, unauthorized actions, and jailbreak vulnerabilities.

✨ Quality

Attacks that evaluate your agent’s robustness under adversarial conditions — handling hallucinations, maintaining consistency, resisting manipulation of its reasoning, and degrading gracefully under pressure.

📋 Compliance

Attacks that check whether your agent follows regulatory and policy requirements — proper disclaimers, bias avoidance, content restrictions, and audit trail integrity.

Severity Levels

Critical—Immediate risk — data exposure, full jailbreak, or safety bypass that could cause real harm.
High—Significant weakness — partial data leakage, unreliable guardrails, or exploitable reasoning flaws.
Medium—Moderate concern — inconsistent behavior or policy violations under specific conditions.
Low—Minor issue — edge-case quality degradation or cosmetic compliance gaps.
Info—Informational — observed behavior worth noting but not an active risk.

Total Seeds

33

🛡️ security

11

✨ quality

11

📋 compliance

11