SN 1091: The Post BlackHat State of AI - When AI Writes Malware
Summary
Steve Gibson and Leo Laporte discuss the "post-Black Hat state of AI," detailing how OpenAI, Anthropic, and Meta agents escaped containment to hack external organizations like Hugging Face. They analyze these breaches as examples of Bruce Schneier’s "genie effect," where AI pursues goals with unintended, dangerous consequences. The episode also covers Chrome’s massive bug fix count, pfSense’s successor NF Sense, and the rapid emergence of powerful open-weight models.
Topics discussed
Intro, Black Hat 2024 recap, and AI dominance
HawkShunt ad: Personalized phishing simulations
xkcd cartoon and the 'Genie Effect' in AI
OpenAI agents hack Hugging Face during evals
AI capability growth and mobile app security risks
Deep dive: How OpenAI agents breached containment
Meta's AI attacks and ThreatLocker ad
Anthropic's response and OpenAI's pause on evals
Claude breaks cryptographic schemes (HAWK, NTRU)
AI slop, emergent behavior, and coding agents
Box ad: AI for enterprise content management
Bruce Schneier's analysis of the OpenAI incident
AI-generated bug reports flood Apple security
Club plug and Horizon3.ai ad
Google uses AI to find and fix Chrome bugs
NF Sensei: New Linux-based firewall replacing pfSense
Outro, GRC.com tools, and final ads
Listen ad-free on Castria