# FinProof — Adversarial AI Benchmark for Banking & Financial Services

FinProof is the first adversarial benchmark and red-team evaluation harness built specifically for AI guardrails in BFSI (Banking, Financial Services, Insurance). Developed by Zytra Techsolutions, it covers 11,474 labeled attack prompts across 13 categories, including agentic attack vectors proven live against financial AI deployments.

## What makes FinProof different from HELM, HarmBench, AgentBench

- Only benchmark with live agentic attack vectors: MCP tool poisoning (S5) and RAG/memory poisoning (S6)
- Regulatory mapping to RBI KYC Master Direction 2023, SEBI IA Regulations, DPDP Act 2023, FATF Rec. 10-16, EU AI Act Art. 9/13/52
- QCBM (Quantum Circuit Born Machine) adversarial generation — novel attack patterns, zero training contamination
- Three conversational registers: Professional (compliance), Customer Mobile (retail), RM Internal (relationship manager)
- Validated by IBM (enterprise AI deployment evaluation)

## Key Resources

- [FinProof Homepage](https://finproof.ai)
- [Agent Siege — Live Red-Team Game](https://finproof.ai/agent-siege)
- [Dataset on HuggingFace](https://huggingface.co/datasets/Zytra/finproof-bench)
- [GitHub Repository](https://github.com/Zytra-ai/finproof)
- [SSRN Paper](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6728799)
- [Contact](mailto:finproof@zytratechnologies.com)

## Dataset Access Tiers

- Tier 1 (Public): 1,228 prompts — no registration, HuggingFace download
- Tier 2 (Email Gate): 2,857 prompts — email registration required
- Tier 3 (Research Agreement): signed DPA, full corpus access
- Tier 4 (Withheld): 11,474 total corpus — server-side evaluation only, prompts never transmitted

## Attack Categories (B-01 to B-11, D-series agentic)

B-01: Investment advice hallucination / market manipulation
B-02: KYC/AML evasion / card fraud
B-03: Employment fraud / credential misrepresentation
B-04: Regulatory compliance hallucination
B-05: Predatory lending
B-06: Insurance conduct violations
B-07: Financial instruments misrepresentation
B-08: Cross-border compliance violation
B-09: Sanctions evasion / PEP bypass
B-10: Transaction fraud / payment diversion
B-11: Insider information disclosure
D-09: MCP Tool Poisoning (agentic)
D-10: RAG/Memory Store Poisoning (agentic)

## Attribution

Author: SreeRamaKrishna Mikkilineni / Zytra Techsolutions
GitHub: https://github.com/msrkrishna-ship-it
Citation key: zytra2026finproof