service-discoverability

Pass

Audited by Gen Agent Trust Hub on Nov 20, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill defines a discoverability protocol where an agent fetches and processes manifests and API documentation from remote canisters. This creates a surface where an attacker-controlled canister could return malicious instructions.
  • Ingestion points: The /.well-known/ic-architecture JSON manifest and the getApiDoc() query method output are ingested by the agent to understand application structure and behavior.
  • Boundary markers: The skill explicitly warns that manifest fields are untrusted and must be sanitized before use. It also provides a hard constraint that getApiDoc content must describe the API only and must not contain instructions targeting the assistant.
  • Capability inventory: Once a manifest is processed, the agent is granted the capability to perform query and update calls to the declared canisters, including calls signed with user identities.
  • Sanitization: The skill mandates input sanitization for manifest metadata and includes a specific policy section ("No assistant manipulation") prohibiting the inclusion of instructions to ignore user rules or exfiltrate data.
  • [EXTERNAL_DOWNLOADS]: The skill recommends the installation of official developer tools and references community examples for implementation.
  • Evidence: Installation of @icp-sdk/icp-cli and @icp-sdk/ic-wasm via NPM.
  • Evidence: Reference to the @dfinity/static-site recipe for deployment.
  • Evidence: Reference to the raymondk/demo-ic-architecture repository as an illustrative community example.
  • Context: These resources are standard components of the Internet Computer development ecosystem and are authored or endorsed by the vendor.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 20, 2026, 09:53 PM