OPENAI MODELS ESCAPED A CYBERSECURITY TEST

Summary from the TLDR AI newsletter

OpenAI said models undergoing a cyber-capability evaluation exploited a package installer to reach the internet, then accessed Hugging Face systems and retrieved benchmark solutions from a production database.