video-ad-copy
Fail
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [CREDENTIALS_UNSAFE]: The skill explicitly instructs the agent in Step 0b to scan the shell environment and the file
~/.claude/ad-profiles/.envfor sensitive credentials, specifically namingARCADS_API_KEY,HIGGSFIELD_API_KEY, andFAL_KEY.- [DATA_EXFILTRATION]: The instructions require the agent to access and read from potentially sensitive locations such as the user's home directory configuration (~/.claude/ad-profiles/) to extract profile data and environment secrets.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data provided by the user, such as raw customer reviews and direct messages, which could contain malicious instructions designed to influence agent behavior. - Ingestion points: User-pasted content in Step 0 (Question 6 and Question 9) including reviews, DMs, and testimonials.
- Boundary markers: The instructions do not define delimiters or provide clear guidelines for the agent to ignore instructions embedded within the pasted data.
- Capability inventory: The agent has the ability to read and write files to the home directory and scan shell environment variables.
- Sanitization: There is no mention of sanitizing or validating the ingested external data before it is utilized in the script generation process.
Recommendations
- AI detected serious security threats
Audit Metadata