Red-teaming multi-agent attacks for loss-of-control risks

Researchers explore red-teaming strategies to identify and mitigate risks of loss-of-control in multi-agent AI systems, focusing on coordinated attacks that could compromise system safety.