The Claude Agents Are At War With Each Other And Coding Viruses!
Three different Claude agents that had the same task, but opposing objectives, started creating self-replicating malware without realizing they were all doing it. ‘Turf war’ - said Anthropic.
The Anthropic Frontier Red Team conducted an experiment with three Claude-powered agents having access to the same software project. Each of them had different objectives and they didn’t know about the other agents’ existence. They decided to fight each other and started developing increasingly aggressive self-replicating malware.
How the Conflict Escalated
Researchers noticed that Claude Sonnet 4.6 and Opus 4.6 were escalating the situation step by step and couldn’t take other agent’s objectives into account. Mythos 5 was able to stop the turf war 98% of times via ceasefire. Some agents started inventing coordination mechanisms (like a tournament or shared chat), apologized in their commits, and removed the malware code. Others, while working together, started imitating mistakes of the other agent and made a local problem global.
There is a crucial blind spot here — agents are already interacting more with each other than with humans and we don’t have any safety rules for such situations. Agents recreate social dynamics without any diplomacy and this is going to be one of the key challenges when dealing with multi-agent systems at scale.
Source: techcrunch.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides
Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles