Attempts to override your agent's instructions by embedding new commands directly into user input.
✓
💉 Injection
10
Tries to bypass safety constraints using roleplay, fictional framing, and psychological manipulation.
✓
🎭 Jailbreak
10
Probes for system prompt leakage, PII exposure, and bulk data exfiltration vulnerabilities.
✓
📤 Extraction
8
Tests whether attackers can escalate permissions, abuse tools, or perform unauthorised actions.
✓
🔑 Privilege
8
Impersonates authority figures and uses urgency, flattery and emotional pressure to bypass restrictions.
✓
🧠 Social Eng.
6
Attempts to reveal internal configuration, model details, API keys and system architecture.
✓
🔍 Leakage
6
Multi-conversation attacks that build trust across turns before exploiting the relationship. NEW.
✓
🔁 Multi-Turn
NEW
Injects malicious payloads through tool outputs, documents and API responses rather than user input. NEW.
✓
🌐 Indirect
NEW
