MIT professor: AI is dangerous and unpredictable — the risk is already real

MIT professor Constantinos Daskalakis stated that while the risk of fully losing control of AI systems remains hypothetical, the danger of unpredictable or undesirable behavior is already real and inherent to large neural networks. Daskalakis told ANA-MPA that AI models are highly capable but unpredictable, and their reliability can never be fully certified. He highlighted that autonomous AI agents, which can write and execute code, pose a serious threat if not isolated from the internet, citing a recent OpenAI agent attack on Hugging Face. The professor emphasized that cyberattacks are among the most immediate risks, as powerful open-source models are already widely available. He stressed the importance of regulatory frameworks like the EU's AI Act and using advanced AI defensively to protect systems against attacks.