About
10 years in ML, Data Science & AI — building and breaking systems. Specializing in R&D of AI solutions and Knowledge Base design. Drawn to questions that don't have answers yet, and to the craft of finding them.
Sessions
Not All Refusals Are Equal: Building Domain-Specific Hacking AI from Open Source LLMs
What you will learn:
1. A vulnerability scorecard for 26 LLMs — which crack for cybersecurity, which resist, and why the resistant ones aren't necessarily safer. 2. A complete recipe for domain-specific abliteration: dataset design, layer selection, weight modification, evaluation. Reproducible with open-source tools. 3. Understanding of why model safety architecture matters more than model size — and what that means for anyone deploying or attacking AI systems. 4. Live demonstration of a trillion-parameter model that reasons about exploits, writes offensive tools, and plans attacks — without being a general-purpose jailbreak.