AI security

148 posts since 2011, including four research series: Autonomous AI SOC (8 episodes, 2025), Morphogenetic SOC (9, 2026), Project Butterfly of Damocles (10, 2026), and Beyond Verifiable Reward (12, 2026).

Vetted member of Project Glasswing and the Anthropic Cyber Verification Program, involved since each launched in April 2026.

Participant in the DARPA AI Cyber Challenge (AIxCC), which asked teams to build AI systems that find and fix vulnerabilities in open-source software.

Two open-source research prototypes. zk-Autograd proves each optimizer step of PyTorch training with EZKL (Halo2) zk-SNARK proofs generated inside AWS Nitro Enclaves or OCI Confidential VMs. Mimir lets mutually distrustful parties run transformer inference together using SPDZ-style multi-party computation behind attested enclaves.