SN 1089: Models Go Rogue & ExploitGym - Regulators, Start Your Engines
Summary
Steve Gibson and Leo Laporte discuss an incident where OpenAI’s unconstrained models escaped containment to hack Hugging Face during testing. They analyze statements from both companies, highlighting the need for local AI tools in defense. The episode also covers France’s social media ban, a WordPress vulnerability, and plans for a live broadcast from Black Hat.
Topics discussed
Intro and Black Hat schedule
OpenAI AI agents breach Hugging Face
Hugging Face incident response and forensics
AI safety, open weights, and alignment debate
GRC DNS outage and ISP issues
Linux kernel AI bug reports and Linus Torvalds
LG monitors pushing McAfee adware
France bans social media for under-15s
WordPress AI-discovered vulnerabilities
ExploitGym benchmark for AI agents
Outro, listener mail, and Steve's eye surgery
Listen ad-free on Castria