claude-agent-sdk
Pass
Audited by Gen Agent Trust Hub on Feb 21, 2026
Risk Level: SAFE
Full Analysis
- Unverifiable Dependencies & Remote Code Execution (SAFE): The skill references various Model Context Protocol (MCP) servers via npx. These are sourced from trusted organizations such as Anthropic and Microsoft, which qualifies for a severity downgrade per the trusted source policy.
- Command Execution (SAFE): The skill enables agents to execute Bash commands, which is its primary purpose. To mitigate risks, it provides comprehensive templates (e.g., templates/hook-agent.py) that demonstrate how to blacklist dangerous commands like 'rm -rf' and 'chmod -R 777'.
- Data Exposure & Exfiltration (SAFE): The documentation provides best practices for managing sensitive credentials via environment variables and includes templates for protecting sensitive files (e.g., .env, .pem, .key) from agent access.
- Privilege Escalation (SAFE): The skill documentation correctly warns against using 'bypassPermissions' mode in non-production environments and does not contain any unauthorized privilege escalation attempts.
Audit Metadata