Defending Against AI-Assisted Invoice Redirection Fraud: Two-Person IBAN & BIC Verification
Why supplier bank-account changes are the #1 financial fraud vector for AP automation agents—and how our verified ERP adapter locks down IBAN mutations.
Key Takeaways
- Supplier bank-account manipulation (invoice redirection fraud) is the highest-severity risk when connecting AI agents to AP inbox and ERP workflows.
- Our SupplierBankAccountAdapter validates ISO 13616 IBAN checksums, BIC format, and account-holder strings before policy evaluation even begins.
- Two independent human approvers must sign off on the exact canonical IBAN digest, followed by an automatic ERP read-back check.
1. Business & threat context
Accounts Payable teams are eager to let AI agents triage vendor emails and update supplier records. Yet business email compromise (BEC) and doctored PDF bank-change letters specifically target automated AP pipelines.
If an agent extracts a fraudulent IBAN from an attachment and updates the vendor master record directly, every subsequent legitimate invoice pays out to the attacker's account. Single-person review or unvalidated string passing fails to catch subtle IBAN swaps.
Zero tolerance on vendor master mutations
Never allow an autonomous agent—or even a single unverified human click—to mutate supplier payout coordinates without two-person SoD and read-back verification.
2. How ActEscrow solves it (Service 01)
In ActEscrow, erp.supplier.bank_account.change is our flagship hardened workflow: strict MOD-97-10 IBAN validation at the gateway, mandatory two-person separation-of-duties approval bound to the canonical digest, short execution leases, and immediate GET read-back against the ERP supplier record.
{
"tool": "erp.supplier.bank_account.change",
"resource": "supplier:SUP-1001",
"parameters": {
"iban": "SE4550000000058398257466",
"bic": "ESSESESS",
"accountHolder": "Nordic Paper AB"
}
}3. Production implementation checklist
When evaluating Manual AP Email Workflows / Custom Scripts alongside a customer-owned enforcement boundary, engineering and finance leaders should verify the following operational controls:
- Validate IBAN MOD-97 checksums and BIC structure at the gateway edge
- Require 2 independent approvers (neither the agent owner nor the delegator)
- Re-read the supplier record from the ERP post-write to confirm the exact IBAN persisted
What IBAN validation does not catch
A clarification that matters for fraud: MOD-97 checksum and BIC format checks catch typos and malformed input, not intent. An attacker's IBAN is a perfectly valid IBAN. A doctored bank-change letter will pass every structural check the adapter performs, so validation is a hygiene layer, not a fraud control on its own. The phrase about this being the top fraud vector above is a common industry concern rather than a statistic we have measured.
The fraud control is the combination around it: two independent approvers who see a clear diff from the old account to the new one, a call-back procedure to the supplier using contact details already on file rather than those in the request, and an evidence record that shows the change was approved and verified.
A hypothetical attack, step by step
Consider a hypothetical mid-sized distributor. A spoofed email from a known supplier attaches a letter asking to move payments from a DE89 IBAN to an SE45 IBAN. The AP agent extracts the new details and submits erp.supplier.bank_account.change for SUP-1001. The adapter validates the structure, the policy requires two approvals, and the action waits.
The first approver sees the diff and phones the supplier on the number in the vendor master, not the number in the email. The supplier knows nothing about a change. The approver rejects with a reason. Nothing reached the ERP, and the evidence chain holds the submission, the rejection and the approver ids, with no bank details stored in evidence.
- Call back on contact data already on file, never on data from the request.
- Treat urgency and new beneficiary countries as reasons to slow down.
- Reject with a reason so the pattern is searchable later.
A note on example data
Example payloads should use clearly fictional supplier names. If you copy the sample above into tests, replace the account holder with an invented name.
Security scope and deployment prerequisites
This is a pilot implementation, not a zero-risk guarantee or compliance certification. Examples describe repository capabilities or deployment recommendations, not proof that a live installation has enabled them.
Every enrolled write fails closed during control-plane outages. Privileged control-plane MFA is mandatory outside Development, but accepted MFA claims do not prove recent authentication or hardware-backed WebAuthn unless the identity provider enforces those properties.
Per-agent deactivation blocks new evaluations and redeems and cancels Pending/Approved actions; it cannot recall an already executing ERP write. Evidence verifies recorded events, not unrecorded ERP activity; independently retained heads and trusted keys are needed to detect truncation or replacement.
Production blockers remain
The static portal still uses sessionStorage tokens (BFF not implemented). Off-host backup signing, immutable backup storage, live IdP MFA/logout testing, restore drills, adapter validation, monitoring, and legal review require deployment-specific work. Anchoring and SIEM forwarding are optional and must be configured and tested. See the repository deployment guide and runbook gap register before using real customer data.
Related deep dives
GRC Automation (Vanta, Drata) vs. Cryptographic Decision Evidence Vaults
Nov 20, 2026 · 3 min readSecurity & AuditPrompt Guardrails (Lakera, NeMo, Llama Guard) vs. Cryptographic Intent Enforcement
Nov 16, 2026 · 3 min readSecurity & AuditInside the Tamper-Evident Decision Evidence Vault: PostgreSQL RLS, Hash Chains & Offline Verification
Nov 10, 2026 · 3 min read