The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to perform discovery probes on external sites.
Evidence: references/checks.md contains several curl commands to check for .well-known endpoints and catalog files (e.g., curl -s -o /dev/null -w "%{http_code}\n" "$SITE/.well-known/ucp").
Evidence: references/checks.md uses python3 -m json.tool to format fetched JSON data and grep to parse robots.txt and sitemap.xml.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection because it ingests and processes untrusted data from external storefronts.
Ingestion points: Content is fetched from $SITE (the target storefront) in references/checks.md, specifically from products.json, robots.txt, and sitemap.xml.
Boundary markers: The SKILL.md file provides specific workflow steps and output templates to constrain the agent's interpretation of the data.
Capability inventory: The skill can execute network requests via curl and process data via python3 and grep as defined in references/checks.md.
Sanitization: There is no explicit sanitization logic, but the skill is strictly limited to reporting status codes and specific metadata, which reduces the risk of malicious instructions influencing the agent's behavior.
[DATA_EXPOSURE_AND_EXFILTRATION]: While the skill performs network operations, these are directed at the target site being audited. There is no evidence of sensitive local file access or exfiltration to unauthorized third-party servers.

commerce-protocol-discovery