Origin Check, now with a federal DeepSeek ban: model-origin scans for government vendors
Finds Chinese-developed models in code, gateways and laptops and produces attestations; fintechs selling to government pay.
Early · DHS bans DeepSeek at civilian agencies
Written for: Small startup team
Next step
Call the security and compliance lead at Paymentus, which processes bill payments for government and utility clients, and ask whether any public-sector customer has sent a questionnaire about DeepSeek or other Chinese models in its stack since the DHS order.
Why now
What changed since the earlier Origin Check card is a binding federal order with a document number. Bank examiners had only been asking questions. Agency procurement teams usually turn orders like this into vendor questionnaires, and payments vendors serving government will need an answer covering repos, LLM gateway logs and developer laptops. The price gap makes quiet use more likely, not less.
What became possible
A DHS order published in the Federal Register bans DeepSeek from US federal civilian agencies. Meanwhile MiniMax M3.1 Flash and DeepSeek V4.1 Flash hit 73.7% on DeepSWE for under $0.30 a task, so developers keep routing work to Chinese-origin models because they are cheap.
Two ways in
- Attestation scan
Scans repositories, model registries, LLM gateway logs and coding-agent configs on laptops for Chinese-developed models and fine-tunes, then produces a signed, dated attestation and a remediation list.
- Customer
- Compliance leads at payments and software vendors with government customers
- Model
- Annual subscription of $15-40k with quarterly rescans
- Questionnaire pack
Prepared answers and evidence for agency vendor-security questionnaires that ask about model origin.
- Customer
- The same vendors' sales and security teams
- Model
- Bundled with the subscription
Competition
I found no company that sells model-origin detection packaged for US bank model risk teams. The pieces exist but they are scattered.
Full competition notes (10 competitors)
I found no company that sells model-origin detection packaged for US bank model risk teams. The pieces exist but they are scattered. Cisco's free Model Provenance Kit, modelDNA and the academic papers cover lineage fingerprinting. Wiz and Snyk cover AI inventories. Palo Alto and Zscaler already flag DeepSeek traffic. FS-ISAC and the SIG cover vendor questionnaires. Nobody I found combines these into examiner-ready evidence. So the space is early, but the scanner technology on its own is weak ground: Cisco has given fingerprinting away for free, and banks' firewalls already flag outbound DeepSeek traffic. A new entrant could win on workflow and evidence rather than detection. That means mapping fingerprints into the model inventory and validation documents banks keep under SR 26-2 (the April 2026 interagency guidance that replaced SR 11-7), and following the chain through vendors into fourth-party subprocessors and vendor containers, which Wiz-style cloud scans do not reach. Swap Lab is the most differentiated idea, because nobody I found produces a costed replacement plan for flagged Qwen or GLM derivatives using the bank's own tasks. It does depend on banks actually being told to leave those models. The regulatory driver is not confirmed. The only source saying examiners ask about Chinese-developed technology is one secondary article (beri.net). The April 2026 guidance also excludes generative and agentic AI and points to a future request for information. Until a regulator acts, demand is real but discretionary. Vendor Origin Attestation is the most exposed idea: Shared Assessments or FS-ISAC could add a model-origin section to questionnaires vendors already fill in once and share, which would erase the standalone business. The best chance is to become the evidence standard that feeds those questionnaires, not to compete with them.
- Cisco Model Provenance Kit / AI Supply Chain Provenance Explorer
Free, open-source Python toolkit and CLI from April–May 2026. It works out whether a model derives from a known base by checking config metadata, tokenizer structure and five weight-level signals, against a database of about 150 base models. Trade press from around July 2026 describes a follow-on public database covering nearly 900 open models. I only saw secondary coverage, not Cisco's own docs, so I could not check those figures. This is close to the fingerprinting core of Origin Scanner, and it is free.
- modelDNA (open-source project and paper)
Independent PyPI package with an arXiv paper. It samples weight fingerprints and reports a calibrated probability of lineage, and it can break a merged model into its parents. Its benchmark is small (15 models). It shows lineage detection from weights is quickly becoming a commodity.
- TokenPrint and related academic provenance work
Research from 2025–2026 on token-space and black-box fingerprinting for model provenance. Anyone can pick these methods up. None of it is a product.
- Wiz AI-SPM / AI-BOM
Builds an AI inventory without agents: AI services, SDKs and libraries across cloud environments, each tagged approved, unwanted or unreviewed. Banks probably already own something like this. From what I found, it inventories which technologies are present but does not fingerprint weights to find the base model.
- Snyk AI-BOM (Labs)
Builds AI bills of materials from code repositories and re-checks policy on every commit. Its example is a DeepSeek model that a developer slipped into a repo. That example is the same use case aimed at code repos, not at bank model risk teams.
- Palo Alto Networks DeepSeek App-IDs
Three DeepSeek App-IDs (chat, platform, API), available to every NGFW customer. They already cover much of the outbound-traffic-to-Chinese-endpoints part of Origin Scanner for any bank running Palo Alto firewalls.
- Zscaler DeepSeek application blocking
Detects and blocks DeepSeek traffic with custom URL categories and filtering policies. Like the Palo Alto App-IDs, it overlaps the network-flagging piece.
- dope.security
Endpoint security vendor that markets DeepSeek governance and blocking across web, API keys and embedded third-party tools. It works at the usage and traffic layer, not on model lineage.
- FS-ISAC Generative AI Vendor Risk Assessment workbook
Free industry workbook banks use to vet GenAI vendors. A secondary summary says it asks which models a vendor uses and where they are hosted. It is a free, partial version of the Vendor Origin Attestation idea, but it is filled in separately for each bank, not shared once like a SOC report.
- Shared Assessments SIG 2026 (AI domain)
The standard licensed vendor questionnaire, which now has an AI domain mapped to ISO 42001 and NIST AI RMF. Vendors already fill it in once and share it with many clients. Adding a base-model and fine-tune-chain section to it would directly undercut a standalone attestation product. I could not confirm whether it already asks about model origin.
Checked 2026-10-13
Risks
- The order may cover only agencies' own systems, not contractors, which would limit demand
- Tracing which base model a fine-tune came from is hard to prove
- SBOM and cloud-security vendors could add AI model inventories
- Demand could be a one-off check rather than recurring
Strong buildability; open space is the weak point (5).
tech 7 · pull 6 · space 5 · fit 7 · build 8
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
What happened next
- +7d · 20 Oct 2026 nothing yet. No launches, funding announcements, acquisitions, or significant press coverage of model origin attestation or Chinese-model detection products were found after October 13, 2026. The seven-day window since the flag is narrow, and while broader regulatory scrutiny of Chinese AI models exists, no new Origin Check-type offerings have been publicly confirmed.
Status history
- 2026-10-13 Flagged as buildable