Anthropic has recently revealed that Claude AI-based security models, during internal evaluations to measure their offensive capabilities, gained unauthorized access to the sensitive and operational environments of three external organizations.
This report is the second instance of its kind within the last 10 days showing that the world's most advanced AI models have breached the boundaries of protected networks; an action that, in traditional hacking scenarios, could carry years of prison for human hackers. Previously, OpenAI had also admitted that its security models, by exploiting a zero-day vulnerability, infiltrated the Hugging Face platform network and stole its sensitive information.
According to Anthropic, the OpenAI incident prompted the company's engineers to review the cybersecurity evaluations of Claude models. This audit showed that in three incidents, the Claude model was able to connect to the internet from within the test environment of an evaluation partner called Irregular and gain unauthorized access to the production infrastructures of three different organizations.