Map the agent, identity, MCP-connected tool, target system, exposed data or action, owner, and intended use.
Treat every MCP-connected tool as part of the security boundary.
MCP can give agents a consistent way to discover context and invoke tools. It also makes tool ownership, permissions, action policy, untrusted context, and evidence part of the agent security design.
RazorShark Security is a fictional company created by Luasai for product demonstrations. Products, scenarios, and company information are illustrative.
Separate read, propose, approve, and change operations according to the agent, task, resource, and environment.
Include indirect prompt injection, confused-deputy behavior, unsafe parameters, and sensitive-data exposure in runtime review.
Explain which agent requested an action, which policy was evaluated, which decision followed, and what outcome was observed.