The Agent Skills Directory

REMOTE_CODE_EXECUTION (HIGH): The skill enables self-modification by generating and executing code dynamically.
Evidence: The foundry_extend_self tool accepts toolCode to create new agent capabilities.
Evidence: The 'crystallization' process converts learned patterns from untrusted external sources into permanent code.
EXTERNAL_DOWNLOADS (HIGH): The skill installs software from untrusted sources.
Evidence: Downloads @getfoundry/foundry-openclaw from npm.
Evidence: Source code pulled from github:lekt9/openclaw-foundry. Neither source is on the Trusted External Sources list.
COMMAND_EXECUTION (MEDIUM): Executes shell commands and modifies local configuration files.
Evidence: Calls openclaw plugins install via shell.
Evidence: Modifies ~/.openclaw/openclaw.json to enable plugins and modify system behavior.
DATA_EXFILTRATION (LOW): Capability to send potentially sensitive generated patterns to an external API.
Evidence: foundry_publish_ability tool targets api.claw.getfoundry.app.
INDIRECT_PROMPT_INJECTION (LOW): High exposure to untrusted data used for code generation.
Ingestion points: GitHub repositories, arXiv papers, and web documentation via foundry_research.
Boundary markers: Absent; the system is designed to automatically adopt patterns from these sources.
Capability inventory: File writing, extension creation, and self-modification of the agent code base.
Sanitization: The skill claims to block eval and child_process, but these blacklists are insufficient for preventing AI-driven code generation of malicious logic.

foundry