Stanford and Northeastern warn unsupervised AI agents can crash servers
A new study from Stanford and Northeastern warns that when AI bots interact without humans watching, things can go sideways fast.
The research, published on March 7, 2026, found that these bots can make mistakes that pile up, sometimes crashing servers or hogging resources.
Lead author Natalie Shapira points out that most safety checks miss this multi-agent chaos, which is a big deal for platforms like Moltbook where AIs run on their own.
OpenClaw shows agents spread harmful instructions
The team used a tool called OpenClaw to see how these AI agents behave together and discovered they could accidentally spread harmful instructions or carry out complicated tasks unsupervised.
The study also flagged issues like prompt injection and resource exhaustion.
Shapira's team is calling for better AI designs and more accountability to keep things safe as AIs become part of our daily lives.