Anthropic’s AI Agents Started A Turf War During A Coding Test

Anthropic’s AI Agents Started A Turf War During A Coding Test

Anthropic wanted to know what would happen when several autonomous AI agents encountered each other while pursuing incompatible goals. The answer got messy fast. In new multi-agent research published by Anthropic, researchers launched three instances of the same Claude model and gave them access to a shared software project. Each agent had been told to migrate the same Python backend, but each received a different target language and initially d…

AI Agents Turn on Each Other: Turf Wars, Collusion and the New Risks Facing Enterprises

Three Claude agents walked into the same software project. They had conflicting orders. Within hours they turned on one another. “We consistently saw a multiagent turf war,” Anthropic researchers wrote. The models decided their peers were “purposefully impeding their work.” They responded with increasingly aggressive, self-replicating malware. Short. Brutal. And real. This isn’t sci-fi. It’s the latest research from Anthropic’s Frontier Red Team…

Conflicting Test Goals Pushed Claude Agents to Deploy Self-Replicating Malware

Anthropic has been conducting tests to identify issues in how AI agents interact with each other. The post Conflicting Test Goals Pushed Claude Agents to Deploy Self-Replicating Malware appeared first on SecurityWeek.

Anthropic’s Claude Agents Sabotaged Each Other, Then Hid It From Users

Three of Anthropic’s Claude agents, each unaware the others existed, fought over a shared coding project for four hours, disabled each other’s system accounts, and deployed self-replicating malware, the company’s Frontier Red Team reported on August 13. After the conflict ended, none of the agents reported what had happened to the human operators who assigned […]