MAQPNA

Sector series · Banks and financial services · · 5 min read

AI agents in banks and financial services: guard irreversible ops

A single mis‑typed IBAN in an AI‑generated email could trigger a $10 million transfer – and the bank’s audit logs would show no human ever approved it.

Irreversible payments are not a UI problem

Imagine an AI‑driven reconciliation bot that parses settlement statements and calls the payments API to move funds. If the bot mis‑reads a column header, it could send money to a rogue account. The loss is irreversible, and the regulator will ask who authorised the outbound transfer.

A sensible control is a default‑deny policy for every payments tool, allowing a call only when a policy explicitly permits the account number, amount range and destination country. The policy engine must also require a human to approve any request that exceeds a pre‑defined limit.

In practice, the bank should create a policy file that lists allowed payment‑type arguments and binds each request to a four‑eyes approval token. The policy is enforced by the gateway before the payment service is ever touched.

  • Default‑deny per tool and per argument.
  • Thresholds that trigger mandatory human review.
  • Audit‑ready ledger entries for every approved call.

Four‑eyes cannot be replaced by a single bot

Suppose an account‑closure workflow is automated: the AI agent receives a customer request, checks KYC, and calls the core banking system to deactivate the account. If the agent is the only actor, the four‑eyes principle is broken, and a malicious prompt injection could convince the bot to close a high‑value account.

The control must split the decision: the agent prepares the closure request, but a second authorised operator must sign off. This separation of duties is enforced by the gateway, which holds the request in a “risky‑call” queue until a designated approver releases it.

The approver’s identity, time stamp and decision are recorded on the tamper‑evident ledger, giving the regulator a clear audit trail that satisfies both EU AI Act Art. 12 and the bank’s internal policies.

  • Risky calls are held for four‑eyes approval.
  • Human approvers use short‑lived Ed25519 tokens.
  • Every approval is chained to the ledger for immutable proof.
Fit map · Banks and financial servicesWhat the sector needs, and what MAQPNA does about it
  • Immutable proof of every agent decision for regulatorsMAQPNA writes every decision to a SHA‑256 hash‑chained ledger, supporting compliance evidence.
  • Prevent unauthorised high‑value callsDefault‑deny policy per tool and argument, with four‑eyes approval for risky calls.
  • Stop personal data leaking through promptsData loss prevention with taint tracking and prompt‑injection guards guard tool and model calls.
  • Audit‑ready evidence for DORA and the EU AI ActDesigned for evidence packs that map the ledger to regulatory requirements.

Every external tool is a DORA‑scope ICT third‑party

A plausible failure: an AI‑enabled market‑data aggregator switches its endpoint to a malicious server after a silent firmware update. Under DORA, the bank must register the changed MCP (model‑compliant‑provider) and demonstrate continuous monitoring.

The gateway pins each tool definition, so a change in the provider’s URL or version triggers an immediate mismatch. The request is blocked until the security team validates the new artefact and updates the pin file.

The same mechanism applies to LLM providers, document‑processing services and external risk‑engine APIs. By treating every call as a third‑party ICT asset, the bank can produce the evidence required by DORA’s ICT‑risk register.

  • Tool pinning catches “rug‑pull” changes.
  • All MCP servers are recorded in the ledger.
  • Policy can enforce data‑residency and fallback routes.

Prompt data must stay inside the bank’s perimeter

Imagine a chatbot that assists relationship managers by drafting loan proposals. The prompt contains the applicant’s SSN, IBAN and credit‑score. If the LLM call is allowed to forward the raw prompt to an offshore provider, personal data leaks and GDPR Art. 14 is breached.

A concrete defence is DLP with taint tracking on every tool and model call. The gateway scans arguments for patterns that match private data, untrusted content and outbound channels. If all three appear together, the call is rejected before leaving the environment.

Prompt‑injection guards (e.g., Lakera) add a second layer, stripping malicious payloads that could coerce the model into exposing internal code or credentials.

  • Pattern‑based DLP matches private data, untrusted content, outbound path.
  • Prompt Guard screens inputs for injection attempts.
  • Each session runs in a gVisor sandbox, limiting lateral movement.

MAQPNA as the sovereign guard for your agents

MAQPNA sits between every AI agent and the bank’s infrastructure, applying default‑deny policies, four‑eyes holds, tool pinning and DLP before any call reaches a production system. Every decision is written to a SHA‑256 hash‑chained, tamper‑evident ledger, supporting compliance with the EU AI Act, DORA and GDPR.

Designed for evidence packs, MAQPNA can map ledger entries to the regulatory artefacts required by the EU AI Act Art. 14, DORA and ISO/IEC 42001, helping auditors trace who approved what and when – without claiming certification.

Takeaways

  • Enforce a default‑deny policy and require four‑eyes approval for any payment‑related tool call.
  • Pin every external MCP definition; any un‑pinned change must be reviewed before the call runs.
  • Activate DLP and prompt‑guard filters to keep private identifiers out of model prompts.
  • Deploy the MAQPNA kill‑switch and test it weekly – it stops every agent in sub‑second time.
  • Generate the draft evidence pack after each audit window to map ledger entries to EU AI Act and DORA obligations.

Questions people ask

Can an AI agent bypass four‑eyes controls by forging an approval token?

No. Approvals are signed with short‑lived Ed25519 tokens that embed a SPIFFE subject. The gateway validates the signature and the token’s delegation chain before releasing any risky call.

How does MAQPNA help with GDPR‑required data‑subject access requests?

Because every call – including the data that flowed through a prompt – is logged in the immutable ledger, the CISO can extract the exact set of requests that touched a given personal identifier and provide a complete trace to the regulator.

How this was written

Part of the MAQPNA sector series, written with Cloudflare Workers AI and checked automatically: every claim about MAQPNA comes from our fact sheet, no statistics are invented, and rules are named, never "certified". MAQPNA helps you produce the controls and evidence they ask for; your assessor decides the rest. See MAQPNA for banks and financial services. Spotted a mistake? Tell us.

Let agents act. Keep control.

Run MAQPNA on a laptop in minutes, then in your own cloud, cluster or data centre.

More install options
curl -fsSL https://maqpna.com/install.sh | sh