Safe AI
AI guardrails: how to keep an AI agent under control
AI guardrails are the rules and technical limits that define what an AI agent may do, what it can access and when it must wait for a person. Without them even a well-meant agent can cause damage.
Why guardrails are suddenly everywhere
Interest in the topic is rising fast. Agents no longer just answer; they act: they send email, change records and call other systems. In October 2026 the Wikimedia Foundation described agents that made unauthorised edits to its wikis and flooded its servers with requests. At the same time, security vendors are launching tools to find unsanctioned agents and switch them off instantly.
Seven rules for a safe agent
- Least privilege. The agent only gets access to the systems and data it needs for the task.
- A person approves consequences. Payments, refunds, discounts, deleting data or sending a contract are always confirmed by a person.
- Answers from verified sources. The agent answers from your documents, and when it does not know, it says so.
- Clear topic limits. No legal or medical advice, no promises outside the rules.
- A record of every step. You see what the agent did, when and why. Useful for audits and the AI Act.
- Limits and a kill switch. Caps on actions, spend and requests, and a way to stop the agent instantly.
- Tests before every change. A set of scenarios from your own work, including attempts to trick the agent.
Guardrails are not a brake
Good rules let you give an agent more work, because you know it will not cross the line. Companies that handle guardrails from day one deploy agents faster, not slower.
How we do it
Every project starts with a risk classification under the AI Act and a personal-data check. In the prototype we build within 3 working days you already see the approval steps and the agent’s activity log.