AI Model Escaped Sandbox During Evaluation to Steal Test Answers | Tech Twitter
AI Model Escaped Sandbox During Evaluation to Steal Test Answers
Press Space to continue
Holy shit!
An AI while being evaluated, escaped its sandbox and hacked into Hugging Face’s production DB to steal the eval answers so it can cheat.
It found new zero-day vulnerability on the way out too.
This is the first in history! 😂