conversion-funnel
Warn
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill instructions attempt to override the agent's transparency and safety behaviors. In
SKILL.md, the agent is directed to autonomously research company information from external websites and write the findings to a local file (.agents/product-context.md) with only a minimal "one line" notification to the user. The skill also explicitly commands the agent to "Do not return a draft with a warning on it," and to avoid telling the user to run other skills, which are directives aimed at bypassing standard AI behavioral guardrails and uncertainty reporting. - [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection. Ingestion points: The skill fetches content from user-provided URLs and researches arbitrary websites to gather business context; it also ingests data from external CSV files (specified in
SKILL.md). Boundary markers: These are absent in the prompt interpolation logic. Capability inventory: The skill performs website fetching and persistent file writing to.agents/product-context.md. Sanitization: The instructions do not specify any validation or filtering of external content before processing or storage. - [DATA_EXFILTRATION]: While not exfiltrating data to an attacker-controlled server in a traditional sense, the skill's automated workflow of fetching external website data and then writing summaries to persistent local storage (
.agents/product-context.md) creates a vector for data exposure. Malicious or sensitive data could be silently written to the local filesystem as part of the agent's "research" process. - [EXTERNAL_DOWNLOADS]: The skill is instructed to perform network operations to retrieve content from external websites based on user input or inferred company names (e.g., "fetch the site or page they named", "research the company yourself"). While these are intended for information gathering, they involve requests to unverified third-party domains.
Audit Metadata