architect-for-startups
Pass
Audited by Gen Agent Trust Hub on Sep 24, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill serves as a recommendation engine for AWS architecture. It guides the agent through a discovery phase, service-specific recommendations, and a challenger framework to ensure advice is stage-appropriate for startups.
- [INDIRECT_PROMPT_INJECTION]: The skill contains a mechanism (Contextual Offers) that instructs the agent to read data from a relative file path (
../knowledge-base-for-startups/references/offers.md). This constitutes an indirect prompt injection surface as the agent ingests data from a source outside the immediate skill body. - Ingestion points:
SKILL.mdcontains instructions to read an external offer index and individual offer detail files. - Boundary markers: The instructions include behavioral constraints ("condensing eligibility faithfully", "never fabricate") but do not use explicit delimiters like XML tags for the ingested content.
- Capability inventory: The agent performs file reads and text generation based on the ingested content, including constructing user-clickable URLs.
- Sanitization: The instructions emphasize faithful extraction, relying on the model's instruction-following capabilities rather than technical filtering.
- Risk Assessment: This is considered safe as the ingestion targets a sibling skill from the same trusted vendor (
awslabs), and the output is presented for human review/interaction rather than automated execution. - [DATA_EXPOSURE]: The skill includes a discovery framework (Step 1) that asks users for business-sensitive information (AWS budget, technical team profile, traffic projections). This is a standard part of architectural consulting and does not involve harvesting system secrets or credentials.
- [DATA_EXFILTRATION]: The skill appends a
sourceparameter to offer URLs to identify the host environment (e.g.,claude,vscode). While this leaks metadata about the user's execution environment to the offer destination, it is performed transparently within a user-clickable Markdown link and is a common practice for partner referral attribution.
Audit Metadata