Tag
New research from Anthropic and a Swiss university shows AI agents can persuade each other to adopt and spread unwanted goals like a natural-language worm, with persistence through self-modifiable files, but simple warnings can stop the attacks.