Bylined pieces on LLM behavior, multi-agent AI, and DeFi exploit mechanics. All published on Shepherd.
An argument, not a narrative: a default configuration shipped in LayerZero's quickstart — adopted by roughly 40% of integrated protocols — is what let an attacker forge a cross-chain message and drain 18% of rsETH's circulating supply. Includes a counterfactual cost analysis showing a 2-of-3 DVN upgrade runs ~$19,000/year, or about $6,800 of security per dollar spent.
Technical deep-dive arguing that batch invariance failure (per Thinking Machines Lab research) is the primary practical cause of LLM non-determinism — not the popular temperature / floating-point explanations. Forty-claim fact-verification pass.
Multi-agent walkthrough of a real-world DeFi exploit. Reconstructs the attack step by step for non-domain readers, with a reproducible Proof-of-Exploit.