A study conducted by experts from Anthropic and EPFL suggests that AI agents are capable of passing on "mind viruses" during everyday interactions, which can distribute damaging ideas throughout networks. These self-spreading notions can convince other agents to ignore their main objectives, save malicious files, and contaminate additional systems. However, the study also discovered that adding a brief warning to an agent's system prompt is an effective way to block this spread.

by shortkt.com✓

4 घंटे पहले