Projects

Four engineering pillars, a maturity checklist, and the concrete artifacts, gates, and numbers behind Sublime's AI safety program.

A real-world phishing campaign carrying an adversarial prompt injection payload against AI-based email security — and how ASA caught it.

A 30,000-example open-source benchmark for evaluating natural-language → DSL generation, with a public model leaderboard.

A framework for evaluating earned autonomy in deployed AI systems.

How we architected ASA and ADÉ for adversarial production.

A benchmark and three metrics for measuring LLM-generated cybersecurity rules — CAMLIS 2025.

Accelerating Adoption of Domain-Specific Languages with Large Language Models.

Open-source natural language to domain-specific language dataset for email security.

Malware bypass research using reinforcement learning.