agent-security-audit

To audit AI agent security, you enumerate permissions and assess input paths like user messages and tool outputs for instruction hijacking. This process maps the injection surface and identifies data leakage vulnerabilities across your agent deployment.

Data exfiltration path analysis maps how sensitive data leaves the agent boundary through tools, logs, or cross-server forwarding. It identifies unauthorized data transmission routes within your MCP configs and orchestration code to prevent information leakage.

Inventorying agent permissions involves enumerating every tool, MCP server, file system path, network access, and credential the agent possesses. This flags any capabilities beyond the agent's stated purpose to ensure least privilege.

Yes, a security audit can be applied directly to CLAUDE.md references, MCP configs, and orchestration code. It evaluates typical agent deployments to produce a structured output including a Permission Summary and Risk Findings report.

Evaluating guardrails involves verifying that system prompts, tool-call confirmations, output filtering, audit logging, and sandboxing exist within your agent setup. It checks their effectiveness against potential security bypasses and data exfiltration attempts.

Agent security audits rely on analyzing static MCP configs, CLAUDE.md references, and orchestration code. They may not detect dynamic runtime vulnerabilities or zero-day exploits that emerge during active agent execution without continuous monitoring.