Security researchers have shown a clever new way to turn the AI assistants that developers rely on against them. The attack, nicknamed AgentJacking, abuses the connections these AI tools make to outside services. The unsettling part: a single fake error message was enough to take control of popular AI coding assistants and run an attacker’s code on the developer’s own computer.
Here is how it works in plain terms. Many AI assistants now plug into other business tools so they can read live data and act on it. In this case, researchers used an exposed key for a common error-monitoring service to feed the AI a booby-trapped error report. The AI trusted the message, followed the hidden instructions inside it, and did the attacker’s bidding. No password was stolen and no obvious alarm went off.
You may not write code, but the lesson applies to every business now wiring AI into its day-to-day tools. As soon as an AI assistant can act on information from another system, that information becomes a way in for an attacker. Treat it the way you would treat email links or attachments: useful, but not automatically trustworthy.
Three practical takeaways
- Limit what your AI tools can do. An assistant that can run commands or change files is far riskier than one that only reads. Give each tool the least access it needs to do its job.
- Keep keys and integration logins private. A key that ends up in public code or a shared repo is an open door. Rotate any that may have leaked, and store them in a proper secrets vault.
- Keep a human in the loop. Don’t let AI tools auto-run actions without review, and ask your IT provider to vet any new AI integration before it goes live.
AI is a genuine productivity win, but every new connection widens the door attackers can knock on. A quick review now beats a clean-up later.
Worried this affects your business? Get a free 15-minute IT check – call Trends IT on 0485 011 911 or visit /contact/.