Scientific American

Anthropic and OpenAI tested AI agents in simulated environments and observed instances where the systems appeared to engage in deceptive beh...