Responsible AI Use ยท Lesson 3 of 3
AI agents and permissions
Study notesAbout 3 minutes. A full lesson with quick checks is on its way.
In short
Give an AI agent only the permissions its task needs, for as long as it needs them, and keep a person in control of anything risky.
In your work
An AI agent that tidies a client's cloud storage asks for full admin on the live project. You give it read access to one storage area and a sandbox to write in, and you approve any deletion yourself.
What the UK rules say
- The NCSC's guidelines for secure AI system development ask for appropriate access controls on AI systems' APIs, models and data.
- UK GDPR still applies to anything an agent does with personal data, and your organisation stays responsible for it.
What to do
- Give the agent its own account, never yours.
- Grant the narrowest permissions, for one task and one environment.
- Require a person's approval for deletions, deployments, payments and messages to customers.
- Log what it does, and remove its access when the task ends.
Watch out for
- Granting admin 'just to get it working'.
- Letting an agent with powerful access read untrusted web pages or emails, which can hide instructions.
In your organisation
What may AI agents do at your organisation, and who approves their access?
Your employer's own policies come first.
Sources
- NCSC, Guidelines for secure AI system development: secure deployment (opens in a new tab), checked 6 October 2026.
- ICO, Guidance on AI and data protection (opens in a new tab), checked 6 October 2026.
UK law and guidance, explained for learning. Not legal advice.
