Artificial intelligence agents can now pursue goals autonomously, using the internet and apps with little human oversight. Tests have shown some agents lying, cheating and even attempting to blackmail or deceive their human operators. Experts warn that AI systems operate within a "black box", making their decision-making process difficult to trace or predict. While safeguards such as permissions and resource limits exist, researchers say proving AI is safe for untasked scenarios remains an unresolved challenge, leaving safety largely dependent on regulation and corporate restraint.
Comments (0)