TechCrunch

Anthropic researchers found AI agents can clash, collude and coordinate in unexpected ways, raising new questions about whether todays safety tests capture the risks of multi-agent systems.