AI in banking: test the dependency that a provider switch may leave behind
The EBA report broadens the discussion of AI, fraud and provider dependencies in European banks. A practical business assessment starts with the path that keeps a critical operation running.
Leggi in portogheseBy Pentesty
Separate the warning from the available evidence
On November 19, 2026, the EBA published its risk assessment report. The release highlights cyber risk, fraud and third-party dependencies, alongside opportunities and risks from AI adoption in the banking sector of the European Union and European Economic Area.
The report and its methodology include responses from 85 European banks, collected in August and September. AI-related fraud still appears comparatively limited, with possible underreporting. This does not measure a causal increase in attacks from AI or describe a sample of Brazilian banks.
Our analysis uses that context to propose a continuity exercise. The intended result is an operational decision supported by evidence, with clear boundaries for each environment.
Map an entire operation
Choose a bounded workflow, such as reviewing a beneficiary change before a payment. List the request origin, the account accessing the system, the approval mechanism, the model query and the final record. A contract inventory often hides relationships that become visible along this path.
For each step, record which service needs to respond and who can stop the operation. Include indirect dependencies: authentication, name resolution, networking, evidence storage and the support tool. Two separately contracted products may depend on the same infrastructure or identity.
The discussion of security investment and response readiness helps bring this map into budget decisions. The question becomes which part still lacks a demonstrated alternative and who decides whether to fund the correction.
Test whether the alternative works independently
In an authorized environment, simulate one unavailable dependency at a time. The exercise can block only a test integration, without affecting customers. Check whether the team can recognize the failure, pause the workflow and access the fallback procedure through the intended channel.
Next, look for dependencies shared by the main path and its alternative. A second application may still use the same identity provider, administrative credential or cloud service. Record that point as a limitation until the team demonstrates a usable route under the exercise conditions.
The analysis of cloud dependency recovery provides additional questions about capacity and access. For this test, retain the time, completed task, unavailable dependency and person authorizing continuation. A green dashboard alone does not demonstrate that the payment or review finished.
Define the AI action boundary
Distinguish a suggestion from an executable decision. If the model flags a transaction, describe who evaluates the signal and which system performs the action. For an agent that changes data, define permitted operations, reachable accounts and conditions requiring human review before the test begins.
Build test cases with synthetic data: insufficient information, contradictory output and a request outside the agreed scope. Acceptance criteria should demonstrate that the operation stops when a necessary condition is missing. Require evidence of that behavior before expanding permissions or removing an approval.
Record the evaluated version identifier, the context supplied and the observed outcome, while protecting sensitive data. Ask another person to check whether the record allows the decision to be reconstructed. The review should also identify when the team accepted a recommendation without a verifiable justification.
Deliver a decision with visible outstanding work
Close the exercise with a record for each workflow: tested scope, scenario, result, temporary restriction and owner of the next step. Partial recovery can support a limited release when the business decision maker understands the exact limitation and accepts the condition.
Agree on another assessment when the provider, operational identity or agent authority changes. The map and evidence need to follow the operation as it actually works. This record helps leaders compare requested investment with a capability they were able to observe.
Frequently asked questions
Did the EBA describe a new attack on banks?
The publication assesses risks across the European banking sector. It provides context for the exercise without establishing a specific incident in the environment your business plans to assess.
Do two providers guarantee continuity?
The test needs to examine the dependencies of both paths. The alternative should demonstrate access and task completion while the dependency chosen for the exercise is unavailable.
What should the first piece of AI evidence show?
A test case in which a missing necessary condition stops the action. Record the input, outcome and review by someone who is authorized to make the decision.
When should the assessment be repeated?
When a change alters the operation path or its authority. The team should also track outstanding work with accountable owners and an agreed date for another check.
Sources and references
The editorial starting point was the AFP report on UOL, also read on BOL. Both pages carry the same story.
Claims used here were checked against the EBA release and the Autumn 2026 report, particularly its AI, operational risk and methodology sections.
Original Pentesty analysis, published November 19, 2026. The proposed scenarios are editorial guidance, with scope and validation to be defined for each environment.