Researchers publish a live guide to hacking service agents

A new security guide looks at how deployed AI customer-service agents can be attacked. The focus is not abstract model safety. It is the agent sitting in front of your customers, handling questions, accessing company information, and potentially taking actions through connected systems.

The practical risk is that an attacker does not need to break into your network in the traditional sense. They may try to manipulate the agent through the conversation itself, pushing it to reveal information, follow instructions it should ignore, or behave outside the boundaries you intended. The researchers outline these types of techniques in their original report.

If you run a customer-facing agent, treat it like any other public-facing application. Be clear about what data it can see, what systems it can access, and which actions require a human check. Test it with deliberately awkward and hostile prompts, not just the clean questions used in a demo. Review conversation logs for repeated attempts to override instructions or extract sensitive details.

This is also a reminder that an AI agent needs operational oversight after launch. This is the kind of thing we build into an AI command centre, giving teams a clearer view of where agents are being used, what they can do, and where the risk sits.