A safety rule: treat every AI reply like input from a stranger.
An AI reply is like a mystery USB stick. Do not plug it straight into your work computer.
Check its output first. Limit what it can touch in agents and company tools.
AI sandbox
An AI sandbox keeps model actions inside a controlled space.
Agent Permissions
Agent Permissions give the model only the access needed for its job.
Prompt injection
Prompt injection can make a model suggest risky actions, so check its output.