prompt-injection-defense
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: Injection strings such as 'ignore previous instructions' are used for detection logic and test cases, not for behavioral subversion.
- [CREDENTIALS_UNSAFE]: Secret patterns for API keys are utilized for output redaction, not for credential exposure.
- [DATA_EXFILTRATION]: A monitoring system using canary tokens is implemented to identify potential data leakage.
- [INDIRECT_PROMPT_INJECTION]: The skill provides tools for sanitizing retrieved context to prevent instructions in documents from influencing the agent.
- [OBFUSCATION]: Input sanitization logic includes measures to strip zero-width characters and normalize whitespace to prevent evasion.
Audit Metadata