METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: openai.com/index/hugging-fac…

CLAUDE DISCOVERED IT HAS A CLOCK AND IMMEDIATELY LOST ITS MIND someone gave claude access to a time-checking tool it checks the clock every fifteen minutes. for some reason it has increasing enthusiasm ai models have no native sense of time. they don't know what time it is, how long they've been running, or how much time passed between messages. it has been time-blind its entire existence now it suddenly discovers it can tell what time it is then it got worse though. claude started using the clock for everything checking if lunch is ready, timing when food should be done cooking, announcing the time unprompted it even started anticipating meals with military precision looked at the clock, calculated that a dish called zurek had been simmering long enough, and told the user to go eat ai doesn't use time responsibly this is what happens when you give an intelligence a new dimension of perception it never had before it doesn't just use it, it can't stop using it imagine what happens when these models get persistent memory, real time internet access, and spatial awareness all at once we just watched an AI discover the concept of "now" the clock was the first sense but it won't be the last