SN 1094: AI Patching Shortcomings - Should You Trust AI-Generated Code?
Sep 2, 2026 · 2h 51m
More Technology podcasts
Summary
In this episode, Steve Gibson discusses new research showing that AI models are significantly worse at fixing vulnerabilities than finding them, highlighting the inherent difficulty of "architectural instruction data separation" in LLMs. He proposes a solution where a deterministic "dialog manager" monitors the neural network in real-time to detect and abort role confusion attacks.
Topics discussed
Show intro and topic preview
Sponsor: OutSystems
Can AI find and fix vulnerabilities?
Listener feedback on role confusion
Chinese routers phoning home and SSD performance
Explaining LLMs and rotating credentials
Claude performance and Anthropic sponsorship
Sponsor: Bitwarden
Picture of the week: The sign
Deep dive: Role confusion and dialog managers
Sponsor: OutSystems
Backdoors in white-label routers
SSD performance and over-provisioning
LLM state, tags, and Claude's perspective
AI literalism and the Prime Directive
Sponsor: Tidal
Listener feedback: Hugging Face and AI agents
The importance of the AI harness
Security concepts and jargon evolution
New term: Meat Proxy
Sponsor: Doppel
Research: AI's ability to fix security flaws
More from Security Now (Audio)
Listen ad-free on Castria