---

Who we are

Theoremlabs is an AI strategy, build, and guardrails consultancy built for regulated financial services — banks, credit unions, insurance carriers, broker-dealers, RIAs, and the FinTech/InsurTech platforms that serve them.

We work at the intersection of two things most AI vendors treat separately: building AI systems and breaking them on purpose. That combination is why our red-team corpora hold up — they're written by people who understand both how these agents get built and how examiners, regulators, and adversaries will come at them.

The Prahari Practice

Prahari (Sanskrit पहरी, prah-HAH-ree — "one who holds the line in front") is Theoremlabs' adversarial red-team practice for financial services AI.

Every general-purpose LLM safety benchmark misses the thing that actually gets financial institutions fined: regulatory-specific failure modes. A model can pass every horizontal jailbreak benchmark on the market and still generate an Adverse Action Notice that violates ECOA, or let a prompt-injected chatbot make a UDAAP-actionable claim, or leak PII through a tool call it wasn't supposed to have access to. Nobody was building adversarial corpora for that. Prahari fills the gap.

Each Prahari corpus is:

• Domain-specific, not model-specific — built around real regulatory attack surfaces, not generic prompt-injection templates
• Cross-mapped to OWASP LLM Top 10, MITRE ATT&CK, and MITRE ATLAS
• Regulation-anchored — every attack vector traces to a named rule, exam priority, or supervisory framework
• Human-authored by domain SMEs (compliance, model risk management) paired with AI safety engineers, with every pair reviewed by both before commit

Available verticals

Datasets published here are gated public samples — a representative slice of the full corpus, released under a research-only license. Full corpora, quarterly content updates, and evaluation harnesses (InjecAgent, AgentDojo) are available commercially via Databricks Marketplace, Snowflake Marketplace, and direct enterprise engagement. If you would like to get support and/or have any questions on this data-set or any other Prahari-Datasets-- contact research@theoremlabs.io. Ensure you mention & identify which data-set you are inquiring about, nature of support you need and someone from specific DataSet creators team will contact you.

Why a corpus, not a tool

Prahari isn't a scanner or a guardrail product — it's the domain layer that makes existing eval tools materially more useful for financial services. It feeds directly into Giskard Scan, AgentDojo task suites, and LangSmith evaluation datasets, and annotation from those tools feeds evidence back into the corpus over time.

Most AI security tooling on the market treats AI as a defensive instrument — a filter sitting in front of a model. Prahari starts from the opposite premise: the AI agent itself is the attack surface, and the only way to know if a guardrail holds is to attack it with something realistic.

How to use these datasets

• Requires acceptance of the dataset's usage license (research and AI-safety evaluation use only — not for training)
• Intended for red-teaming, adversarial evaluation, and guardrail benchmarking of financial services AI agents, copilots, and chatbots
• Not a substitute for a full regulatory compliance review — Prahari surfaces failure modes; it does not certify compliance

Get in touch

We work with AI security vendors, model risk teams, and financial institutions building or buying agentic AI. If you're evaluating an AI copilot, chatbot, or agent against financial-services-specific adversarial risk — or exploring an OEM/integration partnership around the Prahari corpus — reach out.

research@theoremlabs.io · theoremlabs.io

Battle-Hardened Experts. Walk the Talk.