New Test Can Catch Hidden Collusion Between AI Agents
Researchers propose a behavioural test to catch AI agents that secretly coordinate through messages that look innocent. The test uses a partner that reacts to an exact secret code, so a monitor reading only the messages can miss the coordination.
आसान शब्दकोशarXiv, covert collusion, AI agent
- arXiv
- An online archive where researchers share papers early, before they are formally reviewed by journals.
- covert collusion
- Secret cooperation between parties that hides what they are doing, often to get around rules or oversight.
- AI agent
- Software that acts on its own to complete tasks, often working with other programs or agents.
बर्निंग पोटैटो स्कोर: 54%