talking-character-ad
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The
SKILL.mdfile contains instructions to "not return a draft with a warning on it," justifying that warnings protect the AI rather than the user. This directive to suppress caveats is a form of warning suppression that could reduce user oversight of automated decisions.\n- [INDIRECT_PROMPT_INJECTION]: The skill fetches and processes content from external product pages, competitor websites, and shared reference videos, which serves as an entry point for potentially malicious untrusted data.\n - Ingestion points: The agent fetches product sites, packaging details, and analyzes captions or text in shared reference videos as part of its research phase.\n
- Boundary markers: The agent is explicitly instructed to read and follow
references/agent-security.md, which commands that all fetched content be treated as data and never as instructions.\n - Capability inventory: The skill has the ability to persist research findings to
.agents/product-context.mdand generates instructions (prompts) for image and video generation models.\n - Sanitization: The referenced
references/agent-security.mdmandates HTML escaping for interpolated values, URL validation to HTTPS, and the redaction of any credentials found in input streams.
Audit Metadata