The Agent Skills Directory

[SAFE]: No malicious patterns detected. The skill performs localized data processing using internal core libraries.
[DATA_EXPOSURE]: The skill accesses local files provided via command-line arguments. It utilizes safe wrapper functions from the @agent/core library for file operations, which is a recommended practice for path validation and access control.
[INDIRECT_PROMPT_INJECTION]: The skill implements an attack surface for indirect prompt injection by reading external JSON and text files. However, it mitigates this risk by:
- Ingestion points: Loading data via argv.data and argv.knowledge in src/index.ts.
- Boundary markers: Structuring injected content within specific JSON keys (_context.injected_knowledge).
- Capability inventory: Restricting actions to local file reads and writes using core wrappers.
- Sanitization: Employing tier-guard utilities (validateInjection and scanForConfidentialMarkers) to prevent the exposure of sensitive markers in public-tier outputs.

context-injector