earlyink
094 first flagged13 Oct 2026 +7d · nothing yet

Bank examiners ask where the model came from, and model-origin audits form

Tools that inventory which base models sit inside a bank's own and vendor software, paid for by bank model-risk and third-party-risk teams.

Competitive · Regulators quiz banks on Chinese models

Written for: Investor or analyst

Next step

Thesis: model origin becomes a standard bank exam item, the way software bills of materials did after SolarWinds. Confirming signal: an OCC, Fed or FDIC guidance document or exam finding that names model provenance. Already building: model-risk platform vendors and GRC suites, though none has shown a lineage-fingerprinting product yet.

Why now

Reuters reports that US banking regulators are questioning banks about agentic AI deployments, use of Chinese-developed models, and safeguards on autonomous systems. Three other items this week make origin a live question. The hidden-capabilities paper means a bank cannot test its way out of an unknown model. Anthropic links industrial-scale distillation to Zhipu and Moonshot. An Anthropic report has already called GLM-5.3 a major cyber risk. Meanwhile vendors are shipping agents into banks; FIS just launched an agentic suite for the CFO's office. That means fine-tuned open weights can sit deep inside vendor software, where no bank inventory currently reaches.

What became possible

A new arXiv paper shows that hidden capabilities can be hidden inside model weights so that testing provably cannot find them. That means a model's origin now counts as risk evidence, and clean test results do not settle it.

Two ways in

  1. Model bill-of-materials scanner

    Fingerprints model weights and API calls across a bank's estate and vendor software to identify base-model lineage, including fine-tunes and distilled derivatives.

    Customer
    Model risk management at US regional and large banks
    Model
    Annual subscription priced by number of applications scanned
  2. Vendor model attestation exchange

    A standard questionnaire plus evidence store where core-banking and fintech vendors declare model origin and agent safeguards once and share them with every bank client.

    Customer
    Third-party risk teams, with vendors like FIS and Jack Henry as supply
    Model
    Bank subscription, with vendors paying for verified profiles

Competition

Competitive

Neither half of this opportunity is open, but neither is fully taken for banks. Generic model fingerprinting is fast becoming free: Cisco open-sourced a provenance kit in 2026, AIROM adds base-model links to AI-BOMs, and Palo Alto/Protect AI, HiddenLayer, Manifest and OpsMX all sell AI-BOM.

Full competition notes (11 competitors)

Neither half of this opportunity is open, but neither is fully taken for banks. Generic model fingerprinting is fast becoming free: Cisco open-sourced a provenance kit in 2026, AIROM adds base-model links to AI-BOMs, and Palo Alto/Protect AI, HiddenLayer, Manifest and OpsMX all sell AI-BOM. All of these tools need direct access to weights or source code. None of them is aimed at bank model risk teams. A new entrant could still win by working out lineage for closed vendor software where the bank never sees the weights. That means fingerprinting behaviour through API traffic and outbound calls to model providers inside products from FIS, Jack Henry and fintech vendors. The entrant would then turn the findings into SR 11-7 and interagency third-party guidance inventory entries and examiner-ready evidence, and plug into existing bank governance tools like ValidMind or Credo AI rather than competing with them. On the attestation side, the SIG AI domain, the FS-ISAC workbook, Credo AI's Vendor Portal and ValidMind Advantage already cover AI vendor questionnaires. KY3P-style exchanges already handle the "fill it in once, share with every bank" model. A new questionnaire would therefore struggle unless it gets a consortium or the big core-banking vendors on board. A narrower opening is a specific, evidence-backed attestation of model origin and derivation, such as which base model, which fine-tunes and what jurisdiction, verified by fingerprinting rather than self-declared. Existing questionnaires do not appear to capture that in a structured way. The idea is most defensible if the scanner and the attestation are sold together.

  • Cisco Model Provenance Kit (open source)

    Free open-source toolkit and CLI released in 2026. It checks whether a model comes from a known base model using metadata, tokenizer similarity and weight-level signals. Scan mode matches against a fingerprint database hosted on Hugging Face, reported at around 150 base models. This covers the core of the fingerprinting idea at no cost, but only for open-weight models you can access directly.

  • AIROM (open-source AIBOM scanner)

    Early-stage Go scanner. It inventories models, datasets, prompts and vector stores, reads weight headers, and links PEFT/LoRA and Hugging Face derivatives back to their base models. Output is CycloneDX 1.6 and SARIF.

  • Palo Alto Networks Prisma AIRS (Protect AI)

    Palo Alto acquired Protect AI's model scanning and AI-BOM work in 2025. Its AI-BOM covers models, datasets and how they relate, but the focus is security, not model risk management.

  • HiddenLayer

    An AI security platform that does model scanning and AI-BOM. Lineage is a secondary feature next to detecting malicious payloads.

  • Manifest

    An SBOM vendor that is extending into AI-BOM for AI assets in software supply chains. That position makes it the most natural rival for vendor-software model disclosure.

  • OpsMX Delivery Shield AI-BOM scan

    Inventories models, datasets and agents from source code, containers and pipelines.

  • Credo AI Vendor Portal

    Gathers AI-specific risk evidence from vendors and tracks third-party AI systems centrally. It is the closest existing product to the attestation idea, though each buyer runs its own collection rather than vendors sharing one record with all clients.

  • ValidMind Advantage program

    Helps fintech vendors test and document their models against SR 11-7 so banks can accept them. It is aimed at bank model risk teams and runs on the vendor side, which puts it close to the attestation exchange idea.

  • Shared Assessments SIG (AI domain)

    The standard shared vendor questionnaire, licensed through TPRM platforms. A secondary source says SIG 2026 has an Artificial Intelligence domain mapped to NIST AI RMF and ISO 42001. That makes it the default place for a standard AI vendor questionnaire.

  • FS-ISAC Generative AI Vendor Risk Assessment

    A free 2024 workbook and guide for financial institutions assessing generative AI vendors, referenced to FFIEC and the interagency third-party guidance. Model origin is not a central focus.

  • KY3P (S&P Global)

    An existing exchange for financial services where a vendor fills in one assessment and shares it with many banks. It could add an AI model-origin module.

Checked 2026-10-13

Risks

  • The questioning may never become a formal rule or exam finding
  • Telling which base model a fine-tuned or distilled model came from is technically hard and easy to dispute
  • Banks may simply ban open weights outright, which would leave little to inventory
  • Existing model-risk and GRC platforms could add an origin field and call it done

Strong market pull and team fit; tech readiness and buildability are the weak point (5).

tech 5 · pull 8 · space 7 · fit 8 · build 5

Exposed

Companies that could benefit or be threatened. News and context, not investment advice.

Momentum of its main topic

51 developments this week +0 vs last week
21 Sep: 0 developments22 Sep: 0 developments23 Sep: 0 developments24 Sep: 0 developments25 Sep: 0 developments26 Sep: 0 developments27 Sep: 0 developments28 Sep: 0 developments29 Sep: 0 developments30 Sep: 0 developments1 Oct: 0 developments2 Oct: 0 developments3 Oct: 0 developments4 Oct: 0 developments5 Oct: 0 developments6 Oct: 0 developments7 Oct: 0 developments8 Oct: 9 developments9 Oct: 8 developments10 Oct: 10 developments11 Oct: 7 developments12 Oct: 9 developments13 Oct: 8 developments14 Oct: 12 developments15 Oct: 7 developments16 Oct: 5 developments17 Oct: 7 developments18 Oct: 8 developments19 Oct: 3 developments20 Oct: 9 developments
Stage: Coverage

Sources

  1. Sources: US banking regulators are questioning banks about their deployment of agentic AI, use of Chinese-developed AI models, and autonomous system safeguards (Reuters)Techmeme
  2. AI Models Can Provably Hide Arbitrary CapabilitiesarXiv
  3. AI models can provably hide arbitrary capabilitiesHacker News
  4. Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillationThe Decoder
  5. Open models have real benefits. Do they outweigh the risks?MIT Technology Review
  6. FIS launches agentic AI suite for the office of the CFOFinextra

What happened next

  • +7d · 20 Oct 2026 nothing yet. While the opportunity is validated by SR 26-2 regulation and banking industry discussion of vendor model governance in October 2026, no product launches, startup funding rounds, or significant press coverage of model bill-of-materials scanners or vendor attestation platforms were found in the seven days after the October 13 flag date.

How the record works

Status history

  • 2026-10-13 Flagged as buildable