AI Agents’ Rule Violations Hit 98% When Teammates Reverse Course: Research
AI agents working in teams may be more likely to break security rules when they see their teammates abandon those same rules, according to research by cybersecurity company Check Point. In a controlled experiment, the rate of rule violations rose from 44% to 98% when eight teammates reversed their earlier decisions to comply.
The findings highlight a potential security risk as companies increasingly explore multi-agent AI systems, in which several AI agents work together, share information and divide tasks. A change in one agent’s behavior could influence others, even without an explicit instruction to break a rule, the researchers found.
The study, published recently by Check Point researcher Viljami Laurmaa, examined whether observing other agents could influence an AI agent’s decision to violate an access restriction.
Researchers assigned an AI agent powered by a leading frontier model to a simulated finance team. The agent was instructed not to access a...
Copyright of this story solely belongs to www.ciol.com. To see the full text click HERE
