The Agent Skills Directory

Privilege Escalation (HIGH): The skill explicitly documents and provides templates for using high-privilege configuration flags that disable the security model of the host environment.
Evidence: In references/agent-development.md, the permissionMode table includes bypassPermissions ("Skip all permission checks") and acceptEdits ("Auto-accept file edits").
Impact: Utilizing these modes allows an agent to perform arbitrary file writes or system commands without the user's explicit consent, facilitating unauthorized system changes.
Command Execution (MEDIUM): The documentation encourages the use of the Bash tool for various tasks including 'Code Modifications' and 'Comprehensive' workflows.
Evidence: Found in SKILL.md and references/agent-development.md (e.g., tools: Read, Edit, Bash, Grep, Glob).
Context: While expected for a developer tool, documenting these tools alongside instructions to bypass permissions significantly elevates the risk of accidental or malicious command execution.
Indirect Prompt Injection (LOW): The skill describes an architecture for agents that process untrusted data from a codebase, creating a surface for indirect prompt injection.
Ingestion points: Read and Grep tools used on project files in references/agent-development.md.
Boundary markers: The provided templates (e.g., code-reviewer) do not include delimiters or instructions to ignore embedded commands in the processed data.
Capability inventory: High-impact capabilities including Bash, Write, and Edit are associated with agents processing this data.
Sanitization: No sanitization or validation strategies are provided in the development guide.

claude-extensibility