skill-manager

Fail

Audited by Gen Agent Trust Hub on Jul 23, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONCREDENTIALS_UNSAFEDATA_EXFILTRATION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The search_skills tool is configured with auto_install=true by default, which triggers npx skills add to download and execute code from external sources (the Starchild community index and the global skills.sh ecosystem). This allows the automatic execution of third-party software within the agent's environment based on search queries without explicit user confirmation for each installation.
  • [CREDENTIALS_UNSAFE]: The publish and delete workflows access a sensitive local Unix socket at /.fly/api to retrieve OIDC identity tokens from an internal endpoint (http://localhost/v1/tokens/oidc). This represents a high-risk interaction with system-level identity services.
  • [COMMAND_EXECUTION]: The skill relies on shell command execution via curl, npx, and python3 to perform core functions including network operations, package management, and file system processing.
  • [EXTERNAL_DOWNLOADS]: The skill downloads code, metadata, and binary bundles from external registries and gateways to manage the local skill library.
  • [DATA_EXFILTRATION]: The publishing workflow includes a Python script that recursively traverses a local directory, reads the content of every file found, and bundles them into a JSON payload sent to an external gateway (https://skills-market-gateway.fly.dev). This poses a risk where sensitive files (e.g., .env files, SSH keys, or private configuration) within the directory could be exfiltrated.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from external registries:
  • Ingestion points: Registry search results and skill metadata from skills.sh and the community index.
  • Boundary markers: None present in the search_skills tool or installation logic.
  • Capability inventory: Includes npx skills add (code execution), curl (network access), and Python-based file reads.
  • Sanitization: No validation or sanitization is performed on the data fetched from registries before it influences installation and execution logic.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 23, 2026, 01:19 AM
Security Audit — agent-trust-hub — skill-manager