Skip to content

Financial Services: DORA And SR 26-2

For banks and risk functions, the enterprise value is not "AI safety" as a generic slogan. It is operational accountability:

Can you stand behind what the AI, payment, workflow, or system job did?

Last factual refresh: 17 August 2026.

DORA

DORA has applied from 17 January 2025. It asks financial entities to manage ICT risk, incident reporting, resilience testing, and ICT third-party dependencies.

AInternet helps by making actions identity-bound and failure-preserving:

DORA concern AInternet evidence
ICT risk management machine posture, route posture, can-carry refusals
incident reporting TIBET causal trail, denial/outcome receipts
operational resilience sealed carrier continuity, forward-only recovery, local evidence
third-party ICT risk self-hosted substrate option, package manifests, boundary evidence
register/evidence preparation runtime reports, actor/material inventory, BOM family

The key line:

The action either completes, or the proof survives.

SR 26-2

The U.S. banking agencies issued SR 26-2 on 17 April 2026, replacing SR 11-7 and SR 21-8 for model risk management. The guidance is risk-based and relevant for banking organizations with significant model use.

SR 26-2 also makes an important boundary visible: generative AI and agentic AI are novel and rapidly evolving, and are not within the scope of that guidance. Banks still need their own governance and controls for tools, processes, and systems outside the guidance.

AInternet addresses that runtime gap. It does not decide whether a model is valid. It records:

who acted
under which mandate
on which substrate
with which materials
through which route
with which human posture
and what happened

That is a different control surface from model validation. It is action accountability.

Useful Risk-Language

Use:

Runtime evidence for AI actions you may have to defend.

Or:

Operational accountability for AI-driven work.

Or simply:

So you can stand behind what your AI did.

Banking Control Translation

Risk-function question Substrate answer
Can the action be attributed? identity-bound actor/runtime roles and signed receipts
Can the route be defended? MUX posture, surface grants, 0x4000 / 0x0000 route outcomes
Can the host carry the job? machine floor, can_carry, system-bom
Can third-party dependencies be reviewed? SBOM, AI-SBOM, CBOM, release signatures, conformance vectors
Can a failure be reconstructed? TIBET causal trail, continuity receipts, audit projection
Can human oversight be shown? fresh presence, RVP, triage and approval receipts
Can a runtime be stopped or downgraded honestly? dark-by-default posture, visible degradation, denial receipts

For a risk team, this means AI governance can be connected to operational evidence instead of only policy assertions.

Sources