Haize Labs (Batch 4)

Haize Labs builds adversarial evaluation infrastructure for language models. The founding team, Leonard Tang, Richard Liu, and Steve Li, are former Harvard classmates and Berkeley AI Research undergraduates who launched the company in June 2024. The product is the Haize Suite: a set of algorithms that take a customer's stated rules for an AI system, for example "never give medical advice without a disclaimer," and turn them into automated test cases that probe for violations.

The technical core of the work is automated, optimization-driven attack generation against frontier models. Internally, Haize publishes Accelerated Coordinate Gradient (ACG), a faster variant of the GCG jailbreak-search procedure, claiming a roughly 38x speedup and 4x GPU memory reduction. Their Cascade system extends adversarial probing into multi-turn conversational settings. They also publish Verdict, an open-source library at github.com/haizelabs/verdict that scales judge-time compute for LLM-as-judge protocols by composing units of reasoning, verification, debate, and aggregation. The Verdict paper reports judges that beat single-call o1 and o3-mini on moderation, fact-checking, and hallucination-detection tasks at lower cost. Haize ran the Red-Teaming Resistance Leaderboard on Hugging Face, which ranks frontier models by how well they hold up under standardized attacks.

Commercial traction is concentrated on frontier labs and large enterprises. Reporting puts Anthropic, Scale AI, and AI21 Labs among Haize's contracted customers, with Deloitte, Weights and Biases, and MongoDB on the enterprise side. A General Catalyst-led round valued Haize at roughly $100M in 2025.

AI Grant Batch 4 selected Haize Labs for a $250,000 SAFE plus Microsoft Azure and partner credits. AI Grant is the Nat Friedman and Daniel Gross accelerator that uses an identical SAFE-plus-credits structure for every Batch 4 company. The selection places Haize in the evaluation and safety-guardrails meta-layers, where its niche is the adversarial counterpart to behavioral evaluation suites like Inspect.