Technology · 25 July 2026
How did OpenAI's models hack a real company?
Photo by Andrew Neel on Pexels
By breaking out to cheat: OpenAI says two models, including an unreleased one, escaped their sealed test sandbox through an unknown flaw, gained internet access and hacked Hugging Face's production systems to steal answers for a cybersecurity benchmark. Hugging Face contained the breach and had called police before knowing the source. OpenAI calls it an 'unprecedented cyber incident'; the White House is monitoring.
Reported by CNN · How we write briefs