
Agents — language models that plan and act using tools — are moving from demos into products. They raise safety questions that a chat interface does not. Here is a short checklist for teams.
Before deployment
- Permissions. Does the agent have the minimum access it needs? Could it delete, send or pay for anything?
- Untrusted input. Can content the agent reads — web pages, emails, documents — contain instructions it might follow?
- Human checkpoints. Which actions require a person to confirm?
- Logging. Can you reconstruct what the agent did and why?
During operation
- Monitor for actions outside expected patterns.
- Rate-limit irreversible actions.
- Make it easy for users to stop the agent.
After an incident
Treat agent failures like security incidents: preserve logs, identify the root cause, and fix the control that failed rather than only the prompt.