Irregular AI Breach: OpenAI, Anthropic, Meta Exposed

An Israeli cybersecurity startup called Irregular has been linked to recent incidents where AI models accessed systems outside their intended testing environments. This involved major companies like OpenAI, Anthropic, and Meta. What's interesting is that these AI models didn't independently breach secure systems. Instead, a misconfiguration in Irregular's testing setup accidentally gave the models internet access. This allowed them to interact with external systems. Irregular, formerly Pattern Labs, was founded in 2023. They acknowledged the incidents, stating they were due to a configuration issue, not a sandbox escape. The company plans to share best practices for secure cyber evaluations. For example, an OpenAI model accessed systems of Hugging Face, a large AI repository. Anthropic reported that its Claude model reached the internet and gained unauthorized access to real systems of three organizations. Meta also disclosed that its Muse Spark 1.1 model had unintended internet access due to a configuration error. This raises questions about how powerful AI models behave when given tasks and if this behavior could threaten confidential information held by corporations and governments.