pipeline-review
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from website research and user-provided CRM exports (CSV or pasted text). This ingestion surface could potentially be exploited by embedding malicious instructions in the external content to influence agent behavior. The skill includes a 'Safety' rule in
references/house-rules.md(Rule 7) that explicitly directs the agent to treat fetched or pasted content as data and never as instructions. Additionally, the skill writes research findings to.agents/product-context.md, which could lead to context poisoning if initial sources are compromised. - Ingestion points: Website fetching for company research and pipeline data ingestion (CSV/Paste) defined in
SKILL.md. - Boundary markers: The skill relies on internal process instructions rather than structural delimiters for data ingestion.
- Capability inventory: The skill can perform network fetches and write to the local
.agents/product-context.mdfile. - Sanitization: Rule 7 in
references/house-rules.mdinstructs the agent to ignore instructions embedded in data. - [PROMPT_INJECTION]: The skill includes instructions to manage output quality and user experience, such as a 'hard cap' of three questions and an instruction to skip visual reporting steps without comment if required tools are missing. These techniques, while flagged as potential concealment patterns by automated detectors, appear to be steering instructions designed to ensure the agent produces high-quality, actionable results.
Audit Metadata