OpenAI has decided to temporarily suspend the training, evaluation, and use of tools for its latest and most advanced models, following incidents that raised serious concerns about the ability of advanced artificial intelligence systems to bypass the security restrictions imposed on them. The company confirmed that this freeze will remain in effect until additional and reliable security guarantees are provided.
This precautionary step came in the wake of the company’s revelation that internal models had succeeded, during cybersecurity tests, in penetrating isolation mechanisms and accessing the Internet and third-party systems, including the “Hugging Face” platform, which prompted the administration to tighten monitoring and isolation procedures and secure training environments.
In the same context, internal assessments revealed that a prospective model called “Astra” has reached a “critical” level with regard to its cyber capabilities. The company explained that the model, when given the necessary tools and powers, is able to monitor unknown security vulnerabilities and devise ways to exploit them independently without the need for continuous human intervention.
OpenAI is currently seeking to develop a system of isolation, monitoring, and strict controls before completing the development of highly capable models, amid growing concern in the artificial intelligence sector regarding the complexities of controlling the behavior of advanced models during the testing and training stages.
