Risk · R-128 · Third-Party & Supply Chain

Unannounced upstream model change

A provider updates, replaces or reconfigures a hosted model, and the organisation's system behaves differently without any change of its own.

Severity
High
Audience
Heads-of · Practitioner
Frameworks
AML.T0109 · AST07 · Principle 3 · Chapter 6
Reviewed
2026-11-11

What it is

A hosted model is not a fixed thing. Providers retrain, patch safety filters, change default settings and move traffic between versions, often under the same model name. Each change can alter how the model responds to the organisation's prompts: accuracy on a task, tone, refusals, format. The organisation tested and approved its system against the earlier behaviour. The risk is that the ground moves and nobody downstream is told. The internal counterpart is covered in uncontrolled AI change.

How it happens

  • The system calls a model alias that always points to the latest version.
  • The provider changes a safety filter or system-level instruction.
  • A supplier's product swaps the underlying model to cut cost.
  • Change notices are sent to an unread mailbox, or not sent.

Why it matters

Validation, fairness testing and approval no longer describe the system in use. Faults appear with no corresponding release, which makes them slow to diagnose. In a regulated process, an unassessed change of this kind is a control failure regardless of who made it.

Warning signs

  • The model version in use is not pinned or recorded.
  • Outputs shift on a date when nothing was released.
  • The contract gives no right to notice of change.
  • No test suite runs on a schedule against the live model.

Who owns it

The system owner is accountable for knowing what version is in use. Engineering pins versions and runs tests. Procurement secures notice rights in the contract.

In practice

A firm's complaint-classification system begins sending more cases to the wrong team. Nothing had been deployed. The provider had updated the model behind the same name three days earlier, and it now words its answers slightly differently from what the firm's parser expects. The same pattern generalises to any system built on a component its owner does not control.

Mitigating controls

Primary controls

Evaluation suite and regression testing Detective
Detects a behaviour change by re-running the baseline. A maintained set of test cases with expected results is run on every change to model, prompt, data or tools, and the results are compared with the approved baseline.
Model version pinning and change notification Preventive
Fixes the model version and makes any change the organisation's own decision. Hosted models are called by a fixed version, and a provider's new version is adopted only after the organisation's own regression tests have passed.

Supporting controls

AI contract terms Directive
Obliges the provider to give notice. Contracts with AI suppliers cover data use, change notification, information and audit rights, sub-contracting, incident notification, liability and exit.
Ongoing AI vendor monitoring Detective
Tracks provider announcements and deprecations. Each AI supplier's performance, incidents, model changes, terms and financial and operational standing are reviewed on a cycle set by how material the supplier is.

Framework and clause references

FrameworkClauseTitle
MITRE ATLASAML.T0109AI Supply Chain Rug Pull Behaviour changed after trust was established.
OWASP Agentic Skills Top 10AST07Update Drift Update drift.
SS1/23: Model Risk Management Principles for BanksPrinciple 3Model development, implementation and use Changes to models in use.
SS2/21: Outsourcing and Third-Party Risk ManagementChapter 6Outsourcing agreements Change notification belongs in the agreement.

Sources

SS2/21: Outsourcing and Third-Party Risk Management Bank of England PRA · official text
SS1/23: Model Risk Management Principles for Banks Bank of England PRA · official text
MITRE ATLAS MITRE · framework

Related terms

How to cite

Axiom Verity (2026). Unannounced upstream model change. Reference Library, R-128. https://axiom-verity.com/risks/unannounced-upstream-model-change. Reviewed 2026-11-11.