Browse AI Agents & MCP SECTION 8: Safety, Guardrails & Human-in-the-Loop

The Blast Radius Analysis

2 fill-in slots · from The AI Prompt Handbook for AI Agents & MCP

Your details

0/2 filled

The Blast Radius Analysis

Analyse the blast radius of this agent.

Agent: [DESCRIBE]
Tools with side effects: [LIST]
Permissions: [DESCRIBE]
Systems reachable: [LIST]

Work through:
1. The worst realistic outcome if this agent behaves badly
2. The worst outcome if it is prompt-injected by hostile input
3. The worst outcome if it loops on a destructive action
4. For each side-effecting tool: what is the maximum damage from a
   single wrong call, and from 1,000 wrong calls
5. What is reversible, what is expensive to reverse, what is not
   reversible
6. What limits the damage today
7. What should limit the damage but does not
8. The minimum set of controls that would make the worst case
   acceptable

Controls to consider:
- Permission scoping (what it can reach at all)
- Rate and volume limits (how much it can do)
- Approval gates (what needs a human first)
- Dry-run modes (what it can simulate)
- Reversibility (what can be undone, and by whom)
- Circuit breakers (what stops it automatically)
- Monitoring (what detects it in progress)

Rank the controls by protection-per-unit-of-friction.

Highlighted [slots] are still empty. Downloads are plain .txt — paste into any AI assistant.