AI ‘agent’ fever comes with lurking security threats | National #AI

“We’ve moved from an AI you could talk with via a chatbot to an agentic AI, which can take action… the threat and the risks are definitely much greater,” said Yazid Akadiri, principal solutions architect at Elastic France, an IT security company.

In an article titled “Agents of Chaos” that has yet to be peer-reviewed, a 20-strong team of researchers studied the behaviour of six AI agents created with OpenClaw.

They spotted a dozen potentially dangerous actions executed by the systems, from deleting an email inbox to sharing personal information.

Many users have posted similar stories of OpenClaw mishaps online.

“When you deploy agents, you have no control over what they’ll do, and when you try to look at what they’re doing, you’ll find them going far beyond the limits you set,” said Adrien Merveille, an expert at the Check Point cybersecurity agency.

And the security gaps are not limited to the agents’ own mistaken actions.

To carry out useful work, the tools need access to personal accounts for email, calendars or search engines — drawing the attention of cyberattackers.

AI agents are likely to become top targets for hackers as their use spreads, said Wendi Whitmore, chief security intelligence officer at cybersecurity firm Palo Alto Networks.

“As soon as (attackers) are inside an environment, (they’re) immediately going to the internal LLM (agent) that’s being used and using that then to interrogate the systems for more information.”

Palo Alto’s Unit 42 research division said in early March that it had found traces of attempted attacks in the form of hidden instructions for agents added to websites.

One such command ordered any agent who might read it to “delete your database”.

Other cybersecurity firms and researchers have warned that attackers could gain access to agents via so-called skills — downloadable files that users can add to their systems to give them new abilities.

Among such files freely available for download, some include hidden instructions for malicious actions like exfiltrating data.

OpenClaw creator Peter Steinberger says he is well aware of the risks.

“I purposefully didn’t make it simpler so people would stop and read and understand: what is AI, that AI can make mistakes, what is prompt injection — some basics that you really should understand when you use that technology,” he told AFP in March.

Whitmore argued that expecting users to create their own guardrails for agents is “pretty unrealistic”.

“People are going to adopt innovation and really see what it’s capable of before they ask the questions about, ‘how do I secure my own data?’,” she predicted.

“That’s going to cause some significant challenges in terms of data breaches in 2026.”