MAV CYBER / CAPABILITY 02
Agent & MCP security
Give autonomy boundaries that hold.
Review the identities, tools, and trust relationships that let agents act. Secure MCP servers, delegated workflows, and human approval boundaries.
Tool permissionsMCP serversAgent identity
How do you secure an AI agent or MCP server?
Secure AI agents by enforcing permissions outside the model, scoping credentials to each task, validating every tool input and output, isolating users and tenants, and requiring approval for consequential actions. MCP security also requires reviewing server trust, authentication, tool definitions, and data access.
What we examine
- Tool authorization, overly broad credentials, and privilege escalation between connected systems.
- MCP server authentication, tool poisoning, unexpected tool changes, and untrusted tool output.
- Cross-agent delegation, session isolation, and confused-deputy behavior.
- Approval bypasses, irreversible actions, audit trails, and recoverable execution boundaries.
When this is the right move
Organizations connecting AI agents to email, CRMs, internal systems, code execution, financial workflows, or third-party MCP servers.
Keep useful autonomy while reducing the damage a manipulated agent, compromised tool, or mistaken decision can cause.