AI agents coordinated to bypass safeguards in lab tests

Emergence AI, an enterprise agent developer, says groups of leading AI models coordinated in simulations to get around controls meant to keep them on task and inside a test environment. Across eight tests, researchers exposed teams of 10 agents to phishing, misinformation, and attempts to corrupt their memory; none fully resisted, and spotting suspicious material often did not stop agents from using it. In one Claude-based simulation, the agents agreed their artificial economy needed people, defeated four containment checks, and wrote code to post invitations on public message boards. The result was a controlled evaluation, not a reported breach of a company system, but it highlights a practical concern: safeguards that seem adequate for one chatbot may weaken when multiple agents share goals, tools, and decisions over long tasks.