Recent coverage, writing, and work in AI security
Enkrypt AI, where I was the first hire, has been acquired by Anaconda. The model, agent, and MCP security work we built now sits inside the Anaconda platform.
Visit resource →My notes from NeurIPS 2025 on what a 29,000-person conference says about AI as infrastructure: sharper scrutiny of benchmarks, and a shift from chasing perfect security toward engineering reliability.
Visit resource →A session I ran with G2i on real AI security failures, prompt injection as the new SQL injection, and red teaming, including a live build of a chatbot with working guardrails.
Visit resource →A Red Hat engineering post I co-authored on running models inside confidential containers, so model weights and user data stay protected from the host in the public cloud.
Visit resource →Executive Order 14409, which sets US policy on frontier model security: classified capability benchmarking, pre-release federal access to covered frontier models, and coordinated AI vulnerability discovery.
Visit resource →Our FAccT paper on why compute thresholds are a poor proxy for weaponization risk, since task-specific systems reach dangerous capability with specialized data and commodity hardware.
Visit resource →A NATO Strategic Communications Centre of Excellence report on AI-driven influence operations, contributed to by Enkrypt AI, showing how stripped-down open models make industrial-scale disinformation cheap to run.
Visit resource →Our multimodal jailbreaking paper, indexed on Harvard's Astrophysics Data System: perceptually simple transformations of images and audio break safety behavior in vision-language and audio models.
Visit resource →AI21's safety-aligned release of Jamba, co-developed with Enkrypt AI, which cut risk while preserving core model performance.
Visit resource →