Human Expertise Will Always Be The Core Of AI Cybersecurity - Andrew Wesie - ASW #408
Summary
Hosts Mike Shima and John Kinsella discuss the 30th anniversary of the "Smashing the Stack" article and the evolving landscape of vulnerability discovery. They interview Andrew Wiese from Zint on how LLMs and human expertise combine to find complex business logic flaws at scale. The episode also covers the PHP-to-Shell vulnerability, the OpenSSH "SSH Hole" bug, and Anthropic’s recent findings on Linux kernel exploits and distributed AI models.
Topics discussed
Intro: 30th anniversary of buffer overflows
Cloud security and mobile app security stats
Interview: Using LLMs to find business logic vulns
Zint's history and the model vs. harness distinction
Automating pen testing and attacker capabilities
Evaluating models and scaling vulnerability discovery
Finding low-hanging fruit and complex long-standing bugs
Defending with LLMs and the complexity of fixes
Will humans disappear from vuln discovery?
Sponsor: ThreatLocker zero trust enforcement
News: PHP to shell drama and patch distribution delays
Context: PHP bug history and LLM exploitation skills
SSH Hole: Software archaeology and password handling
Curl project: AI-generated vulns and triage transparency
Anthropic: Zero-click RCE in Linux and IP options
Lessons from Anthropic: Tooling and model limitations
Anthropic: Running LLMs on low-power devices
Anthropic: Custom source control for 30k daily merges
The future of memory safety and Rust vs. C
Outro: Language ambiguity and future predictions
Listen ad-free on Castria