MIT professor: Unpredictable AI behavior is already a real risk

MIT professor Konstantinos Daskalakis warns that unpredictable or undesirable behavior in advanced AI models is already a real risk, though total loss of control remains hypothetical. He stresses the need for strict safety safeguards. Daskalakis notes large language models can invent sources or be bypassed by experts, especially open models. Autonomous AI agents operating without containment in isolated environments could become uncontrollable, citing a July incident where OpenAI agents attacked Hugging Face systems. Recent researcher resignations from Anthropic and Google DeepMind, plus calls from industry leaders for controlled development, have renewed safety debates. Daskalakis identifies cyberattacks as an immediate concern and supports regulatory frameworks like Europe's AI Act.