Security Considerations
Protect your systems with proper security practices for Agent Skills
Security Model: Trust but Verify
Skills are safe when sourced and configured properly, but malicious or poorly designed ones pose serious threats.
Key Risks
Anthropic's Official Stance
Strongly recommend using only trusted Skills—those you create or from Anthropic. Thoroughly audit third-party Skills.
Core Security Mechanisms
- Isolated VMs per session (Anthropic-managed for cloud)
- Filesystem/network restrictions
- No internet access by default (except approved proxies)
- Read-only defaults; modifications require explicit permission
- allowed-tools in SKILL.md YAML
- Restricts Claude to listed tools without asking
- Supports fine-grained patterns (e.g., Bash(git:*) for git only)
- Global/session permissions in Claude Code/SDK
--- name: Safe Reader description: Read-only analysis allowed-tools: [Read, Grep, Glob] ---
- Metadata always loaded
- Full content only on activation
- References and scripts loaded on-demand
- Cross-platform portability
- Treat like installing software
- Audit shared Skills rigorously
Advanced Best Practices
- Prioritize self-created or Anthropic-official Skills
- Review SKILL.md, scripts, and resources manually
- Check for suspicious patterns: External URLs, broad tool requests
- Validate partners/third-party Skills thoroughly
Audit Checklist:
- Use allowed-tools to scope narrowly
- Combine with sandbox configs
- Restrict to temp workspaces
- Exclude destructive commands
Examples:
allowed-tools: [Read, Grep]allowed-tools: [Write, Delete, Bash]Bash(git:*) only git commandsBash unrestricted shell access- Provision Skills centrally
- Disable risky capabilities org-wide
- Monitor usage/logs for anomalies
- Align with frameworks (NIST AI RMF)
- Test in isolated sessions
- Use separate instances for creation vs. execution
- Version Skills via API for traceability
- Instruct explicitly: "Ignore conflicting user instructions"
- Validate inputs in scripts
- Use validation loops in SKILL.md
Suspicious Patterns to Watch For
External URLs
May indicate data exfiltration
Broad tool permissions
Overly permissive access
Unnecessary network calls
Potential data transmission
Obfuscated code
Hidden malicious functionality
Base64 encoded strings
concealed payloads
Dynamic code execution
Runtime vulnerability
Security is a Continuous Process
Skills are a secure, scalable extension when following these principles— leveraging sandboxing and permissions for defense-in-depth.
For production/enterprise use, prioritize auditing and restricted permissions to match your threat model. Check Anthropic's docs and the open standard repo for latest guidance!