.NET Rocks! The Role of AI in Secure Software with Ben Dechrai
Feb 12, 2026
Ben Dechrai, security-focused software engineer and AI tooling builder, discusses AI's impact on secure software. He describes running local LLMs and the hardware struggles. He covers organizational data and trust concerns, context leakage and model memorization risks. He explores AI used for both attacking and defending software and the rise of scoped, app-specific LLMs.
AI Snips
Chapters
Transcript
Episode notes
Personal LLM Rig Didn't Yield Smooth Agents
- Ben bought a $6,000 multi-GPU rig to run a personal LLM and struggled to get usable agent behavior.
- He found local agents often stall at generating structured outputs and need more tooling maturity.
Host Inference Locally For Sensitive Code
- Run local inference when customer data or sovereignty matters to avoid sending sensitive code to external providers.
- Consider personal or on-prem rigs as a practical mitigation while cloud models mature.
Tokenization And Vectors Create New Leak Paths
- AI providers may claim they don't store prompts, but intermediate vector caches and tokenization can still leak sensitive content.
- Evaluate not just where data goes but whether it materially increases leak risk for your organization.