The Agent Skills Directory

[PROMPT_INJECTION]: The skill is a data ingestion engine that processes untrusted project files to create repository snapshots, creating a surface for indirect prompt injection.\n
Ingestion points: scripts/generate.py recursively scans the user-provided project directory.\n
Boundary markers: The skill outputs structured JSON, but lacks strong boundaries to prevent the agent from obeying malicious instructions embedded in the project files.\n
Capability inventory: File system access (read/write) and execution of system commands via subprocess.\n
Sanitization: Performs basic comment stripping in scripts/context_engine/analyzers.py, which is insufficient for preventing prompt injection in strings or documentation.\n- [REMOTE_CODE_EXECUTION]: The skill performs dynamic library loading from paths within the target project being analyzed.\n
Evidence: In scripts/context_engine/ts_ast_bridge.js, the code uses require.resolve to find and load the typescript package from the local node_modules of the scanned project. A malicious project could provide a compromised version of this library.\n- [COMMAND_EXECUTION]: The skill executes external system commands using the subprocess module.\n
Evidence: scripts/generate.py and scripts/context_engine/external_context.py execute git commands to collect metadata. scripts/context_engine/analyzers.py executes node to run a local analyzer script.

miloya-codebase