Autonomous agents AI tools

Tools that act on their own: they open files, browse the web, call APIs. The risk is not what the person types, it is what the agent does unsupervised — a central category in both the EU AI Act and ISO 42001.

Critical:

Direct harm is possible: identity fraud, biometric data, or autonomous action without oversight.

High:

Naturally receives sensitive corporate data — source code, contracts, customer records.

Moderate:

Risk depends on what goes in. Requires an explicit classification in your policy.

Low:

Narrow scope and clear vendor governance. A candidate for the allowed list.