Skip to content
← Tools
AI & Agentic Systems

Agent blast-radius calculator

Assume the agent misuses every tool it has. What is the worst outcome, what is irreversible, and do your controls actually constrain it?

Security review does not ask whether the model is good. It asks: assuming it does the wrong thing with every tool it has, what is the worst outcome, and can we undo it? That is answerable from the tool definitions alone.

Controls shrink the radius; they do not shrink the tool set. The only tool that cannot be misused is one the agent does not have — which is why we argue the work sits in scoping the layer beneath the model, not in prompt engineering above it.

Categorisation is inferred from tool names, descriptions and MCP safety annotations. It is a prompt for the conversation a reviewer should have, not a substitute for reading the implementations — a tool called get_report that quietly writes an audit row will be scored as read-only here. Runs in your browser.