Report security issues via GitHub Security Advisories (preferred) or email. Do not open public issues for security vulnerabilities.

OpenAgentLayer hooks are guardrails, not walls. They catch common mistakes and enforce best practices but are not a security boundary against determined adversaries.

• Sophisticated prompt injection attacks
• Adversarial code in untrusted repositories
• Supply chain attacks via dependencies

For production use, combine hooks with:

1. OS-level sandboxing -- macOS Seatbelt, Linux bubblewrap, or Docker containers
2. Permission deny rules -- settings template blocks access to ~/.ssh, ~/.aws, credentials
3. Code review -- @nemesis should review changes before committing; use /cca:git-workflow for Git hygiene
4. Minimal permissions -- only grant tools each agent needs (enforced via agent frontmatter)

Use Claude Code's built-in /sandbox command, or run with a custom Seatbelt profile restricting filesystem and network access to the project directory.

bwrap --ro-bind / / --dev /dev --proc /proc --tmpfs /tmp \
  --bind "$PROJECT_DIR" "$PROJECT_DIR" \
  --unshare-net \
  claude

Mount only the project directory. Do not mount ~/.ssh, ~/.aws, or other credential stores.

Always update to the latest Claude Code version.

Before installing third-party skills or plugins:

1. Read the SKILL.md or plugin source code
2. Check for hidden instructions or obfuscated commands
3. Verify the author's reputation and repository history
4. Prefer plugins from verified marketplaces with audit trails