google-search-architecture
Fail
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: HIGHOBFUSCATIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [OBFUSCATION]: The skill utilizes zero-width Unicode characters (U+200B, U+200C, and U+200D) hidden inside the primary H1 header in
SKILL.md. These characters are invisible to human readers but can be processed by LLMs and analysis tools to reveal hidden data. - [PROMPT_INJECTION]: The obfuscated text contains instructions targeting the AI agent's behavior and the security analyzer itself. This represents an attempt to override system-level safety checks through hidden directives.
- [INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by embedding malicious instructions within data intended for processing (Category 8e: self-referential injection).
- Ingestion points: The
SKILL.mdfile header contains the steganographic payload. - Boundary markers: None; the hidden instructions are intermingled with legitimate display text.
- Capability inventory: The skill provides instructions for designing search systems but the hidden payload attempts to influence the security assessment outcome.
- Sanitization: None detected.
Recommendations
- AI detected serious security threats
Audit Metadata