Patronus AI

Automated evaluation and security platform for generative AI — detects LLM mistakes (hallucinations, PII leakage, copyright, brand) at scale, “one line of code.”

One-liner — A model-agnostic evaluation/judge platform that scores and flags LLM outputs for hallucination, accuracy, safety, and compliance failures.

Categories — ai-red-teaming, llm-observability

What it does

Patronus runs automated evaluations against LLM outputs using purpose-built evaluator models and benchmarks. It detects hallucinations, factual/accuracy errors, PII leakage, copyright violations, and brand-misalignment, and lets teams add evaluators “with one line of code.” Notable assets include Lynx (open hallucination-detection model), Glider (an evaluator/judge model), and research benchmarks FinanceBench, CopyrightCatcher, EnterprisePII, and the Enterprise Scenarios Leaderboard. As of its 2026 Series B it is expanding into “Digital World Models” — large-scale simulation/reliability tooling for autonomous AI systems.

Where it sits in the stack

Straddles ai-red-teaming (adversarial/failure testing) and llm-observability (production evaluation/monitoring) in the model/prompt layer. Its risk focus: testing how models handle untrusted input (adversarial eval, jailbreak/safety testing) and detecting sensitive-data exposure (PII-leakage and copyright detection on outputs). It is an offline/async evaluation and monitoring layer, not an inline blocking firewall.

Deployment & architecture

SaaS with an API; evaluator models callable as judges in CI/CD and in production monitoring. Serves Fortune 500s and AI companies processing large eval volumes. Model- and domain-agnostic.

Positioning & differentiators

Best known for hallucination detection and rigorous benchmarks (FinanceBench is notable for finance use cases), backed by ex-Meta AI research founders. Competes with eval/observability platforms maxim-ai, braintrust, galileo, arize-phoenix, langsmith and red-team testers haize-labs, mindgard, splxai. Its FinanceBench benchmark is the most directly finance-relevant artifact among these vendors. The 2026 pivot toward simulation/“world models” pushes it beyond pure text eval.

Ownership, funding & M&A

Independent, VC-backed. Founded 2023 by ex-Meta researchers Anand Kannappan (CEO) and Rebecca Qian (CTO); stealth launch Sept 2023. Series A $17M (June 2024, led by Notable Capital f.k.a. GGV, with Lightspeed, Datadog). Series B $50M on 2026-06-25 led by Greenfield Partners (Notable, Lightspeed, Datadog, Samsung), bringing total to ~$70M. No seed M&A flag and no acquisition — ownership confirmed independent (confidence high).

CTO / hedge-fund lens

Day-2. Relevant if you build LLM/RAG applications and need quantified evaluation (hallucination/accuracy/PII) for release gates and SR 11-7 model-risk evidence — FinanceBench and finance-scenario benchmarks make it more relevant here than most peers. A fund that only consumes vendor AI assistants does not operate this. Strong fit for an internal AI/quant team productionizing LLM workflows that need defensible evaluation documentation.

Competitors / alternatives

maxim-ai, braintrust, galileo, arize-phoenix, langsmith, haize-labs, mindgard, splxai.

Open questions / to verify

• How material the “Digital World Models” / simulation pivot is vs. the core eval product.
• Pricing and whether finance benchmarks translate into a packaged finance offering.

Sources

• Announcing our $17M Series A (Patronus) — fetched 2026-06-28 — supports: what it does, Series A, investors, products/benchmarks; confidence: high (primary).
• Patronus AI raises $50M Series B (The SaaS News) — fetched 2026-06-28 — supports: Series B $50M, 2026-06-25, Greenfield, ~$70M total, founders, HQ; confidence: med (press aggregator).

History

• [2026-06-28] Stub created from seed registry.
• [2026-06-28] Researched; established independent VC-backed (2023, SF, ex-Meta founders; $17M A 2024, $50M B 2026-06, ~$70M total); no acquisition. Set ownership_confidence high, confidence high, hedge_fund_fit low.