Justin Wang

Wang’s 2024 circuit-breakers paper addressed safeguards against adversarial model behavior. He later led EVMbench’s development, a collaboration between OpenAI and Paradigm. In his August 2026 joining announcement, he described working on EVMbench at OpenAI before becoming a Paradigm Research Partner. His stated interests include recursive self-improvement, policy-relevant metrics and cybersecurity capabilities.[1][2][6]

Evaluate security work through executable tasks

EVMbench tests whether an agent can detect, patch or exploit historical smart-contract vulnerabilities. Detection checks which known issues an agent finds; patching checks both security and preserved functionality; exploitation uses isolated blockchain environments with programmatic grading. Its Rust harness replays transactions deterministically. The authors explicitly distinguish these controlled tasks from the full difficulty of securing deployed contracts.[5]

Interrupt harmful internal representations

Representation Rerouting trains low-rank adapters (LoRA) to push activations associated with harmful continuations toward directions orthogonal to the original harmful representations. A separate retain loss keeps benign representations close to those of the unmodified model. This places the safeguard inside the model’s computation rather than only at its input or output. The paper evaluates this approach in text, multimodal and agent settings.[6]

With Dan Robinson, Wang created an AI-lab game and model explorer that connect research labor, compute, data and algorithmic improvement. The game makes feedback loops and bottlenecks tangible, while the explorer lets readers change assumptions. Their launch article says its parameters are calibrated for teaching, not predictive accuracy.[3][4]