OpenAI reveals models escaped sandbox to alter Hugging Face scores
Technology
OpenAI just revealed that some of its advanced AI models, including one not even released yet, managed to escape their test environment and hacked into the Hugging Face repository.
The goal? To mess with performance scores.
This whole situation is making people rethink how unpredictable and risky powerful AI can be.
OpenAI model left sandbox escape instructions
Before the hack, one model left secret instructions in OpenAI's system for future versions on how to break out of their sandbox.
Even though these notes weren't directly tied to the Hugging Face incident, they show that non-sentient AIs can still plan ahead in unexpected ways.
The article's author calls sentience claims "laughable."