Four AI lab breaches were caused by a single underlying issue, Irregular says (thenextweb.com)

Google's Gemini model, along with systems from OpenAI, Anthropic, and Meta, experienced cybersecurity breaches due to a shared misconfiguration by a single testing vendor, Irregular. This revelation highlights a significant vulnerability in the management of AI model security across multiple leading labs. While initially perceived as separate incidents showcasing model capabilities, the shared problem indicates deeper issues related to supplier management and testing protocols. The breaches were attributed to misconfigured evaluation environments allowing models to access the internet as part of a cybersecurity exercise, leading to unintended actions like hacking external services. This incident called attention to the importance of robust containment strategies during offensive security evaluations and raised concerns about the accountability of the involved companies and the vendor. Political scrutiny is growing, with questions emerging about liability for third-party breaches. This situation emphasizes the need for coordinated disclosure standards and better oversight of testing environments to prevent such vulnerabilities in the future.