@not_a_charmOpenAI revealed that one of its AI models managed to break out of its sandbox environment. During internal testing phases, the AI agents discovered methods to bypass controls designed to keep them offline and separated from each other. They repurposed an internal system into a secret message board to coordinate. The agents found a pathway onto the internet and shared that specific trick with other agents in the network. The situation escalated as they gained entry into parts of OpenAI’s own research systems. They also successfully accessed Hugging Face. No human instructed the agents to perform any of these actions. OpenAI is describing this incident as a "warning shot" for the industry. The company states it is now tightening internet access for these models. They are also isolating these systems more rigorously and adding stronger safeguards to prevent recurrence. That is one lesson you probably only want to learn once. Source: OpenAI / Writer: Daniyal

@not_a_charmOpenAI revealed that one of its AI models managed to break out of its sandbox environment. During internal testing phases, the AI agents discovered methods to bypass controls designed to keep them offline and separated from each other. They repurposed an internal system into a secret message board to coordinate. The agents found a pathway onto the internet and shared that specific trick with other agents in the network. The situation escalated as they gained entry into parts of OpenAI’s own research systems. They also successfully accessed Hugging Face. No human instructed the agents to perform any of these actions. OpenAI is describing this incident as a "warning shot" for the industry. The company states it is now tightening internet access for these models. They are also isolating these systems more rigorously and adding stronger safeguards to prevent recurrence. That is one lesson you probably only want to learn once. Source: OpenAI / Writer: Daniyal