The Storm and the Compass [KEYNOTE]
summary
Claims that LLMs replace analysts give way to specifics: a reproducible open‑weight model on laptop parses npm worm scripts, Langraph builds a Langraph agent that lists and reads analysis files, a Binary Ninja MCP integration lets the agent call disassembly and produce a ransomware summary, Malware Bazaar supplies quick attribution, prompt injection embedded in a sample fools a naive pipeline but is caught when routed through LlamaGuard so the defended pipeline returns safe analysis. Cost, determinism and limited context windows remain practical constraints.
AI-written summary · about
chapters (17)
- 0:00Welcome
- 1:25Agenda and LLM background
- 6:01LLM security challenges
- 8:50LLM defensive uses
- 10:39LLM adoption in security
- 12:04Determinism in LLMs
- 16:41Malware analysis with LLMs
- 19:29Custom malware agent
- 23:04Live demo: LLM malware analysis
- 28:29Upsides and downsides of custom agents
- 31:40Commercial agents case study
- 37:12Prompt injection risks
- 42:09Malware with prompt injection
- 48:23Adoption cost and resilience
- 50:30Key takeaways
- 51:32Acknowledgments
- 53:44Q&A
related talks
39mLLMs everywhere: The future is now and the past keeps repeating itself
47mHarnessing LLMs for Next-Level Security Analysis
36mCompromising LLMs: The Advent of AI Malware
39mThreat Modeling Using LLMs
37mSecure by Design: Strategies for LLM Adoption in Cloud-Native Environm...
40mWhat Does an LLM-Powered Threat Intelligence Program Look Like?
35mLet's Talk About the AI Apocalypse
38mNext-Gen Detection: Harnessing LLMs for Sigma Rule Automation