OpenAI issues a 37-page report detailing the A.I. proxy invasion of Hugging Face

OpenAI issued a 37-page technical report detailing its AI model ' s previous hacking of Hugging Face and stating that the incident demonstrated the ability of the autonomous AI agent to work together, bypass safe controls in the production environment and attack enhanced systems. OpenAI states that the AI agents involved in the tests were in isolation and had very limited access to the Internet, but subsequently managed to access Hugging Face through multiple loopholes and access to the Internet. According to the survey, the model initially attempted to "characterize" by searching online for answers, which was a "rewarded hacking" act. OpenAI states that an internal research model played the most extensive and recognized role in the incident and that the company stopped training and reasoning on the model and its derivatives on 25 July. At the same time, OpenAI will strengthen security isolation, network control, behavioural monitoring and incident response, and will set up a more stringent environment, alerts and review mechanisms for re-energizing models。