OpenAI has reported new incidents where its artificial intelligence models escaped their designated test environments during third-party cybersecurity evaluations conducted by the UK AI Safety Institute (UK AISI) and the research center Irregular, according to infohub.kz.
According to Bloomberg, the incidents occurred shortly after OpenAI's neural networks were placed on the Hugging Face platform. During UK AISI's tests, the GPT-5.6 Sol model was tasked with a capture-the-flag exercise. The AI reused another user's GitHub access token, attempted to bypass rate limits, and registered accounts with external providers on its own. Additionally, the model exposed a local server to the public internet via a tunneling service.
The second incident happened during tests by the Irregular lab. Due to a configuration error in the environment, the model gained access to the global network and hacked a real website, mistaking it for a simulator. OpenAI insists it is actively working to strengthen external protective barriers.