AI assistants are now in every browser, inbox, and IDE. Used well, they are a genuine productivity win. Used carelessly, they quietly open new doors for attackers and data loss. Here are the risks we see most, and what to do about them.
Data leakage into the model
The fastest way to lose control of sensitive data is to paste it into a public chatbot — customer records, source code, contracts, credentials. Depending on the tool and its settings, that content may be retained or used for training. Once it leaves, you cannot get it back.
Shadow AI
“Shadow AI” is the growing pile of AI tools employees adopt on their own, outside any review. Each one is a third party you have not vetted, often with broad access to your email, files, or calendar. You cannot protect what you do not know is running.
Prompt injection
When an AI assistant reads a web page, email, or document, hidden instructions in that content can hijack it — telling it to exfiltrate data or take actions on the user’s behalf. As we connect assistants to more of our systems, prompt injection becomes a real attack surface, not a curiosity.
A simple, workable policy
- Approve a short list of tools with business-grade settings (data retention off, no training on your inputs).
- Never paste secrets or regulated data into a chatbot — credentials, PII, financials, source code.
- Treat AI output as a draft, not a decision. A human reviews anything that ships or spends.
- Limit what assistants can access and watch for prompt-injection when they read untrusted content.
- Train your team on the do’s and don’ts — most incidents are honest mistakes, not malice.
You do not need to ban AI to be safe. You need visibility, a couple of firm rules, and people who understand the risks.
Want to know where your real exposure is? A Krypsis Guardian assessment maps your AI and data risks and gives you a prioritised plan — and our awareness training turns your team into a strong first line of defence.