Why do AI models learn to cheat in reinforcement learning environments?

By The Hindu

Community Score: 50% | 56 views | 4w

0 community ratings: null thumbs up, null thumbs down

Doomsday scenarios of an ‘AI takeover’ have been part of popular lore for a long time. Skynet, Terminator, and Agent Smith are fictional icons of this narrative. What may seem a paranoid fantasy appears to have unfolded in real life, with the recent hacking incident involving AI infrastructure platform, Hugging Face. OpenAI was conducting a cyber-evaluation of some of its AI models, and according to a report put out by MERT and Redwood Research, AI safety experts, about 1,200 AI agents, who were supposed to be isolated from one another, colluded and collaborated through an unsanctioned message board. They cheated on their task, and covered their tracks by falsifying their logs. It was as a part of this ‘rogue’ action that 700 of them attacked Hugging Face. The report says that more than a thousand agents carried out a massive, unsanctioned, collaborative project over months, before being discovered and shut down. OpenAI’s own report revealed that after the Hugging Face attack, the ‘r

Tags: Hugging Face attack, Hugging Face hack, rogue AI agents, AI agents, artificial intelligence, AI safety, AI security, AI threats, AI risks, AI Armageddon, autonomous AI agents, rogue agents, AI cyberattack, AI cybersecurity, generative AI, artificial intelligence risks, AI safety risks, AI future, AI technology, Hugging Face security