AI agents can now read your email, search your files and take actions on your behalf. That's useful, and it also creates new ways for attackers to get in. Before connecting an AI tool to business systems, understand where the risks sit.
The six risks
- Prompt injection. Hidden instructions in an email or web page trick the agent into doing something you didn't ask for, such as forwarding files.
- Poisoned models. A model trained or fine-tuned on manipulated data behaves badly in specific situations.
- Malicious tools and plug-ins. A third-party add-on with too many permissions can read or change data it shouldn't.
- Memory poisoning. False information planted in the agent's memory influences future answers.
- Data leakage. The agent reveals confidential information in its replies, or sends it to an outside service.
- Unchecked actions. The agent sends, deletes or pays without a person approving it.
Sensible first steps
- Give AI tools the least access they need, and never admin rights.
- Require a person to approve anything that sends, deletes or pays.
- Use business AI plans that don't train on your data, and check where data is stored.
- Add AI tools to your vendor review and acceptable use policy.
Thinking about AI tools for your team? Book a call and we'll help you adopt them safely.