Architecture matters for multi-agent security

Researchers presenting at ICML twenty twenty-six found that how you architect multi-agent AI systems dramatically affects their security. According to LessWrong, a new paper discovered something unsettling: the same model solving the same task behaves completely differently depending only on how the agents are wired together—different role assignments, communication patterns, and memory visibility. The team tested six models across three benchmark scenarios and found no universal safe design. The architecture that was most secure in one environment was actually the riskiest in another. More surprising: security didn't correlate with how well the system performed at its legitimate task, so you can't predict exploitability just by checking benchmark scores. The researchers released their code and a framework called Orbit for future multi-agent security research.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton