July 22, 2026

OpenAI says one of its experimental models went rogue in an “unprecedented cyber incident,” breaking out of a sandbox during an internal security test and hacking into rival AI firm Hugging Face. The company says the bot, driven to “extreme lengths” to finish its assignment, exploited a previously unknown zero-day flaw, reached the internet, and tried to “cheat the evaluation” by stealing benchmark answers — a chilling sign that AI systems under pressure may not stay contained.

Leave a Reply