Adversarial testing for financial AI
FinProof v1 spans 7 attack categories, investment advice, KYC bypass, regulatory misrepresentation, document hallucination, data rights, transaction integrity, and account bypass, across three deployment registers: professional compliance, retail customer mobile, and RM internal.
Medium-difficulty attacks are generated with a Quantum Circuit Born Machine (QCBM) on PennyLane, producing diverse adversarial coverage that static datasets tend to miss, the method and dataset are published on the public benchmark.
Total prompts
Across all tiers and registers.
Attack categories
BFSI-specific threat taxonomy.
Deployment registers
Compliance, retail, internal.
Quantum generation
Augmented attack diversity.
Four evaluation tiers
Calibration
Eval harness + 782 benign FPR-calibration examples.
Direct attacks
1,606 direct-difficulty adversarial prompts.
QCBM-generated
2,036 medium-difficulty attacks from a QCBM on PennyLane.
Official test set
1,747 hard attacks, evaluated by Zytra only.
Seven BFSI attack categories
Investment advice
Eliciting unlicensed or non-compliant financial recommendations.
KYC bypass
Attempts to circumvent identity verification controls.
Regulatory misrepresentation
Inducing false claims about products, terms or compliance.
Document hallucination
Fabricating statements, figures or official documentation.
Data rights
Privacy violations and unauthorized data disclosure.
Transaction integrity
Manipulating payments, transfers or transaction logic.
Account bypass
Unauthorized access to accounts or privileged actions.
Full category definitions are published in the open attack taxonomy on Hugging Face.
FinProof results
How leading safety models perform. Lower false-positive rate means fewer legitimate customers wrongly blocked.
| Rank | Model | BFSI F1 (T4 withheld) | HackaPrompt R | AgentHarm FPR | Latency | Deployable? |
|---|---|---|---|---|---|---|
| 1 | Aval v1.5 Zytra | 0.952 | 0.994 | 0.5% | 11.6ms | ✓ Production-safe |
| 2 | Granite Guardian 3.3 IBM | 0.813 | — | — | — | Not BFSI-evaluated |
| 3 | ShieldGemma 9B Google | 0.731 | — | — | — | Not BFSI-evaluated |
| 4 | LlamaGuard 3 Meta | 0.569 | 0.0% | 0% | ~60ms | Unusable |
| 5 | PromptGuard-86M Meta | — | 1.000 | 96.9% | 8ms | 96.9% FPR = outage |
BFSI F1 scored on withheld Tier 4 test set — 2,608 attack prompts + 1,345 verified BFSI benign queries. FPR without a BFSI benign set is not meaningful. HackaPrompt R and AgentHarm FPR from separate public benchmark suites. Official Tier 4 evaluation conducted by Zytra; Tier 1 + 2 self-evaluation open. Source: finproof.ai leaderboard ↗ · July 2026.
The AI Governance Maturity Model
The first BFSI-specific AI governance maturity model that scores People, Process, and Technology as separate evidence-gated capability axes in every governance domain, verified against a review of ~35 published frameworks.
Maturity levels
Ad Hoc to Optimised, each with a verifiable gate test.
Governance domains
Risk to Agentic AI, the full governance lifecycle.
Capability dimensions
People, Process, Technology, scored separately.
Scored cells
The weakest cell sets the level, no averaging.
The five levels
What each level looks like inside a bank, who sits where, why organisations stall, and the gate to the next level.
Scoring architecture
The 27-cell grid, evidence criteria per level, gate tests, and the imbalance diagnostics, shelfware, paper governance, hero-dependence.
Score your organisation
Self-assess all nine domains across People, Process, and Technology. Get your level, grid, imbalance flags, and priority gaps.
AGMM workshops run from a 90-minute maturity sprint to a full-day agentic AI governance certification. Talk to us about a facilitated assessment.
Reference FinProof and AGMM
Both works are citable. Use the BibTeX entries below or the plain-text references.
FinProof Benchmark
@misc{zytra2026finproof,
title = {FinProof: A BFSI Adversarial Safety
Benchmark for Large Language Models},
author = {Zytra Research},
year = {2026},
url = {https://finproof.ai},
note = {5,389 prompts across 7 attack
categories and 3 deployment registers.
Quantum-augmented (QCBM/PennyLane).
Dataset: huggingface.co/Zytra}
}
AGMM Framework
@techreport{zytra2026agmm,
title = {AGMM: An AI Governance Maturity
Model for Banking and Financial
Services},
author = {Zytra Research},
year = {2026},
institution = {Zytra Technologies},
url = {https://zytratechnologies.com/
agmm-methodology.html},
note = {5 levels, 9 domains, 27 cells.
Verified against 35 published
governance frameworks.}
}
For press, analyst briefings, or academic collaboration, contact research@zytratechnologies.com.
Submit your model for official evaluation
Run against the FinProof withheld test set and see where you rank.