Model Origin Scan: find Chinese-developed models in a fintech's stack

The basic scanning is already done. Cisco's open-source AI BOM, AIROM, aibom-scanner, Snyk, Wiz and Checkmarx all find models, SDK imports and weight files and produce CycloneDX or SPDX bills of materials. AIROM also outputs SARIF for GitHub Code Scanning, so a CI gate is a small step from there. Lineage is also being worked on, notably by Cisco's Model Provenance Kit and modelDNA, which fingerprint weights instead of trusting Hugging Face's self-reported base_model field. In none of the sources I found was country of origin a headline feature. I found no scanner that ships a maintained list of signatures for Chinese providers (the DashScope, Zhipu and Moonshot SDKs, their endpoints, and the Qwen and DeepSeek fine-tune families), and no product that turns a scan into evidence formatted for sponsor-bank vendor-risk questionnaires. The opening is narrow. A new entrant would have to be the best at maintaining the origin list and following lineage chains, ideally by building on or contributing to Cisco's AI BOM or AIROM rather than writing a new parser. The paid part would be the dated questionnaire pack and the exception workflow for BaaS fintechs and AI vendors selling to banks. Three risks stand out. Cisco, Snyk or Wiz could add an origin field and a blocklist in one release. I found no OCC, FDIC or Fed guidance that requires origin attestation, so demand depends on individual sponsor banks asking. And base_model metadata is unreliable, so the report's lineage claims need weight fingerprinting behind them to hold up in an audit.

• Cisco AI BOM (open source, also in Cisco AI Defense)

  Open-source scanner from a large vendor. It inventories models, agents, tools and prompts in codebases and container images, and outputs JSON, CycloneDX and SPDX. Trade press says it covers several languages, but the PyPI page only describes Python parsing. It is the closest match to the origin-scan CLI, but nothing I found mentions a country-of-origin or Chinese-lineage feature.

• Cisco Model Provenance Kit

  Open-source toolkit released in summer 2026. It checks whether two models share an origin by comparing architecture metadata, tokenizers and weights. This is a stronger form of lineage detection than reading the self-reported base_model field on Hugging Face.

• AIROM

  Apache-2.0 Go scanner, at version 0.3.x and early-stage by its own description. It finds hosted-model API calls, model-ID strings and local GGUF, safetensors and ONNX weight headers, then outputs CycloneDX 1.6 with file and line evidence plus SARIF for GitHub Code Scanning. Much of the CLI and CI idea is already here, minus origin labels.

• aibom-scanner

  Apache-2.0 Python tool that detects AI SDKs across 30+ providers and produces an AI bill of materials with compliance risk findings. It focuses on SDK imports, not weight files or lineage.

• Snyk AI-BOM scanner

  Experimental Snyk Labs feature that builds an AI bill of materials from code through Snyk's hosted API. Snyk already sells to fintech engineering teams.

• Wiz AI-SPM (AI-BOM and model scanning)

  Commercial cloud-side tool. It finds hosted models (PyTorch, TensorFlow, Hugging Face-sourced) on VMs and in storage buckets and scans them for malicious behavior. The source doesn't mention any country or publisher policy.

• Checkmarx LLM Scanner

  Commercial scanner that inspects model packages and produces an AI bill of materials, sold to AppSec buyers.

• OpsMX AI-BOM Scan

  Builds an AI bill of materials from a codebase, container or cloud environment as part of a software delivery and security platform.

• modelDNA

  Open-source tool that infers model lineage from the weights alone. It points out that Hugging Face's base_model field is optional and unverified, and says over 60% of Hub models have no documented parent.

• OWASP GenAI AIBOM Generator

  Free community tool that turns a Hugging Face model ID into a CycloneDX AI bill of materials and reports documentation gaps. It works on single models, not on a whole repository.

• dope.security (DeepSeek governance)

  Network-layer shadow-AI tool that markets blocking or governing DeepSeek. It watches traffic, not code, but it competes for the same 'prove we don't use Chinese AI' budget.

• Security Scientist Chinese AI Model Due Diligence Checklist

  Excel template and toolkit for assessing the risk of Chinese AI models. It is a manual alternative to the questionnaire pack, aimed at compliance leads.

Checked 2026-10-13