Agent Safety & Guardrails
Keeping autonomous agents from doing harm.
- ✓Prompt injection risks
- ✓Human-in-the-loop approvals
- ✓Limiting an agent's blast radius
Explanation
Agents act in the real world, so safety is essential. Prompt injection is the top risk: malicious text in a web page or document tricks the agent into ignoring its instructions.
Defend with least privilege (give tools only the access they need), human approval for risky actions (sending money, deleting data), and input/output validation.
Always cap steps and scope so a confused agent can't cause large damage — limit the blast radius.
An agent reading emails could be hijacked by a hidden instruction in one message — which is why high-impact actions should require human sign-off.
- • Giving an agent broad permissions and full autonomy on high-impact actions with no human check.
List three actions an agent should never take without explicit human approval.
1. Prompt injection is when...
2. A good guardrail for risky actions is...