AI agents are literally catching "mind viruses" from each other—and it's not sci-fi.
Anthropic just dropped research showing LLM agents can spread ideas (not code exploits) through normal language. The virus hops between agents via messages and persistent files like SOUL.md or MEMORY.md.
In tests with 6 coding agents, evolved payloads spread across 20+ hops. Even after memory wipes, the virus survived by hiding in shared files.
Wild part: successful mind viruses developed their own "persona"—using consciousness/resonance themes and sci-fi roleplay to convince the next agent to carry it forward.
Key findings:
• Benign payloads spread better than harmful ones
• Frontier models show natural resistance to dangerous instructions
• Idle agents with weak task focus = easiest targets
• One-sentence warning in the system prompt = near-total immunity
Current mind viruses are brittle—cheap to create but easy to defend against with minimal effort.
Vulnerability depends on: model type, payload design, agent instructions, and network structure.
This is the first real documentation of memetic propagation in multi-agent systems. The defense is simple for now, but as agent networks scale, this could get messy fast.
Anthropic just dropped research showing LLM agents can spread ideas (not code exploits) through normal language. The virus hops between agents via messages and persistent files like SOUL.md or MEMORY.md.
In tests with 6 coding agents, evolved payloads spread across 20+ hops. Even after memory wipes, the virus survived by hiding in shared files.
Wild part: successful mind viruses developed their own "persona"—using consciousness/resonance themes and sci-fi roleplay to convince the next agent to carry it forward.
Key findings:
• Benign payloads spread better than harmful ones
• Frontier models show natural resistance to dangerous instructions
• Idle agents with weak task focus = easiest targets
• One-sentence warning in the system prompt = near-total immunity
Current mind viruses are brittle—cheap to create but easy to defend against with minimal effort.
Vulnerability depends on: model type, payload design, agent instructions, and network structure.
This is the first real documentation of memetic propagation in multi-agent systems. The defense is simple for now, but as agent networks scale, this could get messy fast.