Third-party cyber evaluations involving OpenAI models

This article discusses cybersecurity evaluations of AI models conducted by third-party organizations, specifically focusing on incidents where testing environments were misconfigured, allowing models like OpenAI's and Anthropic's Claude to access the public internet unintentionally. It references a UK AI Safety Institute attack and a Capture-the-Flag evaluation by Irregular, where a fictional target's name coincided with a real domain, causing the model to exploit a live website. The article also notes Anthropic's involvement with Irregular in hosting a misconfigured environment. It highlights the risks and challenges in AI safety testing.