How do you control what AI agents can do?
In a supported coding agent, your policy decides each tool call before it runs.
What happens when an AI agent acts without approval?
It spends, sends or changes things with the person’s own access. In 2025 a newspaper columnist asked a shopping agent from a large AI provider to find cheap eggs. Without his approval, it used his saved card to buy a dozen for $31.43. The agent was meant to ask before any purchase, and its maker said it “fell short of its safeguards”. The only check sat inside the agent, so when the agent skipped it, nothing else stood in the way.
What can an AI agent do on a staff device?
Anything the person running it can do. One prompt can reach every tool the agent can call, with that person’s access.
- Files and folders. Files deleted.
- Shell commands. Commands no one approved.
- Databases. via MCP. Records copied out.
- Email and calendar. Mail sent as the person.
- Business apps. via MCP. Records changed.
- The web. Hidden instructions followed.
What do security agencies say about putting agents to work?
That an agent you can’t watch or stop isn’t ready. One national cyber security agency puts it plainly. “If you cannot understand, monitor or contain an agent’s actions, it is not ready for deployment.”
Another government’s framework for agentic AI tells organizations to enforce human approval “through system-level controls where possible”, because guardrails inside the prompt “may be bypassed or ‘forgotten’”. Its list of actions that need approval includes deleting data, sending communications and making payments.
How do you prove AI agents followed your policy?
With an audit trail, a record of each action that shows who ran the agent, on which device, what it tried and what your policy decided.
Enforce decides each call. A denied call never runs.
Audit keeps the record. Each captured call goes on the record under keys you hold, and a denied call is its own entry.
Contain takes one person out. If an agent goes rogue, contain that person or device while everyone else keeps working. That helps you meet MANAGE 2.4 in NIST’s AI RMF 1.0, which asks for a way to disengage AI that strays from its intended use.
How do you put AI agents to work?
All postsStart small, and widen with evidence. The same cyber agency advises “tightly bounded pilots using clearly defined tasks”, and building confidence “before you expand the scope”.
Let AI agents work, and show your policy held.



