jailbreak-prevention
Pass
Audited by Gen Agent Trust Hub on Jun 29, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill documentation and configuration templates include common jailbreak strings (e.g., 'ignore previous instructions', 'DAN') and adversarial patterns. These are used strictly as examples for blocklist configuration and detection logic, aligning with the skill's primary defensive purpose.
- [PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection as it is designed to ingest and process untrusted user data.
- Ingestion points: Untrusted user input enters the agent context during the ingestion phase defined in SKILL.md.
- Boundary markers: The skill explicitly recommends isolating user input within XML tags (e.g., ) and adopting structured formats like ChatML to mitigate risks.
- Capability inventory: The skill does not implement or request access to high-risk tools or system-level capabilities; its logic is focused on filtering and validation.
- Sanitization: The skill advocates for normalization (NFKC), regex-based static analysis, and semantic similarity checks using embeddings to filter input before it reaches the core processing stage.
Audit Metadata