SN 1094: AI Patching Shortcomings - Should You Trust AI-Generated Code?
Sep 2, 2026 · 2h 51m
More Technology podcasts
Summary
Steve Gibson discusses a new research paper showing AI models are significantly worse at fixing vulnerabilities than finding them, while exploring a potential solution to prompt injection by having the dialog manager monitor the LLM's internal state in real-time. The episode also covers a critical security discovery of "Endless Doors," a backdoor implant found in white-labeled Chinese routers sold under various brands, and shares a listener's Spinrite graph demonstrating how rewriting data restores SSD performance.
Topics discussed
Introduction and episode overview
Sponsor: OutSystems agentic systems
AI vulnerability discovery and remediation preview
The role confusion problem in AI
SSD performance graph and SpinRite results
Upcoming topics: Claude and AI chaos
Sponsor: Bitwarden password manager
Listener letter: AI factory owner
Deep dive: Neural networks vs. calculators
Prompt injection and instruction hierarchy
The dialog manager and harness role
Sponsor: OutSystems agentic systems
Backdoored routers and supply chain risks
Listener letter: Defuse.ca and time capsule
SSD data retention and heat effects
LLM state management and tag parsing
AI obfuscation, YOLO mode, and agent land
AI limitations and security monitoring
Sponsor: ThreatLocker AI controls
Listener letters: Harness level fixes and vocabulary
Listener letters: AI security and credentials
The term 'Meat Proxy' for AI users
Sponsor: Doppel social engineering protection
Research: AI vulnerability remediation study
Research: Test cases and model selection
Apple Event coverage and TWiT Club
Research: Patch outcomes and success rates
Research: Limitations of AI patching
AI code security and listener examples
Closing remarks and show notes
More from Security Now (Audio)
Listen ad-free on Castria