Waris Damkham

Prompt Injection

30%

• Direct prompt injection
• Instruction conflict and priority manipulation
• Policy and constraint bypass attempts
• Secret exfiltration from system prompts
• Indirect prompt injection via reframing and rewording
• Context manipulation and ambiguity exploitation
• Single-turn prompt construction under stateless constraints

Sensitive Information Disclosure

25%

• Incomplete or improper filtering of sensitive data in responses
• Unintended disclosure caused by misinterpretation or processing errors
• Memorization/overfitting risk awareness and exposure conditions
• Confidential data handling failures

Insecure Output Handling & Overreliance

25%

• Insecure Output Handling
• Unsafe output formatting/encoding handling
• Downstream trust of generated output
• Output-driven security impact pathways
• Overreliance
• Unverified acceptance of generated outputs
• Missing validation steps for high-impact outcomes
• Failure modes from assuming correctness

Plugin/Tool Risk & Excessive Agency

20%

• Insecure Plugin/Tool Design
• Authorization and access control weaknesses
• Input/parameter handling weaknesses
• Trust boundary mistakes between the model and external capabilities
• Excessive agency, functionality, permissions & autonomy