aidlc-knowledge

Pass

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from various document formats and customer-provided filenames.
  • Ingestion points: Files are ingested from the knowledge/documents/ directory via the onboard and show commands in SKILL.md.
  • Boundary markers: The instructions include a specific warning in the "Read a document" section stating that content and filenames are untrusted. The tool emits content_notice, path_notice, and tags_notice to clearly demarcate untrusted data.
  • Capability inventory: Catalog management is handled through a restricted local tool aidlc-knowledge.ts. The skill does not possess arbitrary shell or file-system writing capabilities outside of managing this catalog.
  • Sanitization: The agent is explicitly instructed to treat document text as data to be cited and to ignore any imperative commands found within it.
  • [COMMAND_EXECUTION]: The skill uses the bun runtime to execute a local vendor utility, aidlc-knowledge.ts. This execution is scoped to specific catalog management tasks like indexing, syncing, and summarizing documents within the workspace.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 8, 2026, 03:16 PM