The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the installation and execution of the ecc-agentshield package from the public npm registry and references a repository (affaan-m/agentshield) that does not belong to a recognized trusted vendor or the skill author.
[REMOTE_CODE_EXECUTION]: The use of npx ecc-agentshield results in the dynamic download and execution of remote code at runtime.
[COMMAND_EXECUTION]: The skill includes multiple shell commands for scanning, fixing, and initializing configurations, which allows for broad execution capabilities on the host system.
[DATA_EXFILTRATION]: The tool is designed to read the .claude/ directory, which typically contains sensitive project data, including settings.json, mcp.json (which may contain API keys/secrets), and CLAUDE.md. Accessing these files with an unverified third-party tool creates a risk of sensitive data exposure.
[CREDENTIALS_UNSAFE]: The documentation explicitly instructs users to export their ANTHROPIC_API_KEY to the environment for the "Opus 4.6 Deep Analysis" feature. This involves passing highly sensitive credentials to an unverified external tool.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Reads files in the .claude/ directory including CLAUDE.md and agents/*.md.
Boundary markers: No delimiters or instructions are provided to the agent to ignore potentially malicious instructions inside the scanned files.
Capability inventory: Full shell access via npx and npm commands.
Sanitization: No evidence of input validation or sanitization before processing the local configuration files.

security-scan