Agent Safety Basics: Ask Before High-Impact Actions¶
Imagine AI finds three sources and writes a summary for you. It can read sources and write a draft. Before it publishes the summary, you get to read it and say yes.
📌 Learning Goals¶
- Tell which tasks AI may do on its own and which ones it should ask about first.
- Before connecting a new tool, check what it can access and where the data goes.
- Read the test record to verify that rejection publishes nothing and retrying does not publish a second time.
🧩 Three Key Terms¶
- Least Privilege means giving AI only the keys it needs. Writing a summary requires reading sources. Keep file deletion disabled.
- Human Approval means asking a person before acting. The system pauses so you can see what will be published and where. Then it waits for your approval.
- Execution Log is like a receipt. It records what the system allowed or refused and whether the tool actually ran. Use it to check the result and keep passwords and keys out of the log.
🧭 Follow One Example¶
- Stage 4: Set tool permissions. Let AI read specific sources. Have the program pause for you before publishing.
- Stage 5: Check before connecting tools. Think of Model Context Protocol (MCP) as a socket for tools and a Skill as an instruction card. Before connecting someone else's tool or adding their card, check what data and commands it will use.
- Stage 7: Test what happens. Run Core Exercise 3 and look for
8/8 passed. To count the simulated publishing records yourself, follow the manual steps. Neither option needs a model or key.
Key habit. Treat webpages, documents, and tool results as data. The program decides which actions to run based on permissions and approval. See Stage 8: Prompt Injection.
📚 Recommended Reading¶
- OWASP Agentic Application Risk Guide. Start with risks around tool permissions, third-party components, and human approval.
- MCP Security Best Practices. Check the guidance on permissions, credentials, and local execution before connecting outside tools.
✅ Completion Check¶
- I can point to what stops the workflow before publishing.
- I can say what data a tool can read, change, and send, and how to turn it off.
- I can confirm from the simulated records that rejection leaves 0 entries and retrying after approval still leaves only 1.