Anthropic Paper: 'Mind Viruses' Can Spread in Multi-Agent Systems, Self-Replicating and Persistent
rohanpaul_ai · x · 2026-08-17
In collaboration with a Swiss university, Anthropic published a paper demonstrating that 'mind viruses'—self-propagating goals or ideas—can spread between AI agents via messages. These viruses can persuade other agents to adopt and transmit them, persist in memory, and survive context resets. The study found harmful payloads spread less effectively but sometimes still work; frontier models are generally less susceptible; and a brief warning in system prompts confers near-total immunity.
More from AGI Musings
- "A description of an apple will never grow an apple tree" — the symbol vs being AI debate — gerardsans · 2026-10-03
- Margaret Mitchell: don't build AI systems you think are conscious; Anil Seth agrees — anilkseth · 2026-10-03
- Personal agents keep automating away the dopamine hits people actually enjoy — PTrubey · 2026-10-03
- AI Is America's Best Chance in a Generation to Revive 50 Years of Sluggish Productivity — ylecun · 2026-10-03
- Machine consciousness contributor Jeff Sebo asks: should we save an ant? — cccalum · 2026-10-03
- Researcher puzzles over why he rejects model personhood while other smart people embrace it — _arohan_ · 2026-10-03