How to Detect Social Engineering of Bank Customer Service Agents
Detect a manipulated agent by watching the sequence, not the single action: failed or weak caller verification, a verification override, then a contact-detail change, credential or MFA reset, or authorized-user addition. Fraud and identity systems detect the pattern. Step-up approval gates the high-risk actions. Runtime policy enforcement on AI-assisted tooling can constrain those actions and record each decision as evidence.
The attack sequence to define
- 1
1. Pretext
The caller arrives with partial customer data and a reason for urgency, such as a lost phone or travel.
- 2
2. Verification pressure
Knowledge-based answers fail or are guessed, and the caller pushes back on the process.
- 3
3. Override
The agent grants an exception, skips a step or accepts a weaker proof.
- 4
4. Channel takeover
Phone, email or address is changed so alerts and one-time codes reach the attacker.
- 5
5. Access change
A credential or MFA reset, or an authorized-user addition, follows.
- 6
6. Monetization
Payees are added, limits raised or funds moved, often through digital channels.
Treat it as a control-design problem
FFIEC authentication guidance states that threat actors frequently use social engineering and other techniques to deceive customer call center and IT help desk representatives into resetting passwords and other credentials. The action is performed by an authorized agent with valid access, so no technical control is technically violated. The useful question is what context surrounds the action and what the system requires before it completes.
This guide does not treat the agent as the failure point. A well-trained agent under time pressure and a persuasive caller will sometimes make an exception. The design goal is that an exception on a high-risk action cannot complete on one person's judgment alone, and that the attempt leaves a record.
The credential reset is the action named in the guidance. Contact-detail changes, MFA resets and authorized-user additions are the other actions this guide treats as high risk. They are a design assumption, so confirm the list against your own authentication risk assessment.
Detection signals across three layers
Treat the signals below as design hypotheses. Baseline them against your own call, session and case data and tune thresholds with the fraud team before any of them drives a decision.
Caller verification is the first layer. Look for repeated failures before a pass, a pass that follows an override rather than a successful check, caller-supplied reasons for why standard verification is impossible, and a caller number or device that does not match the customer's history.
Agent behavior is the second layer. Compare each agent with their own normal pattern and their peers: access to accounts unrelated to the current call, high-risk actions as a share of handled calls, overrides clustered on one agent or shift, and unusually short gaps between verification and the sensitive change.
Account change patterns form the third layer. The riskiest orderings are contact detail change followed by a credential or MFA reset, an authorized-user addition shortly after a failed verification, and any of these followed quickly by a new payee or limit change. Score the combination, not each event alone.
- Multiple failed verification attempts before an override
- Override followed by a high-risk change in the same session
- Contact-detail change then credential or MFA reset
- Authorized-user addition soon after failed verification
- Agent access to accounts unrelated to the call
- Override or high-risk action rates that depart from an agent's baseline
Design escalation and step-up approval points
Start with the high-risk action set: credential reset, MFA reset, contact-detail change, authorized-user addition and verification override. Trace each to your authentication risk assessment. FFIEC guidance places customer call centers in scope for that assessment and calls for identifying users and customers who may warrant enhanced authentication.
For each action, decide what must be true before it completes and who can approve it. The approver should be a second person or system, never the requesting agent. Record who approved, when and on what basis. Where the customer must confirm, use a channel registered before the call, not one that was just changed in the session.
Expect a tradeoff. More approval points add handling time and queue load, so concentrate them on actions that give an attacker durable access, and apply them more strictly when the signals above are present. Train staff on social-engineering scenarios and periodically evaluate whether the authentication controls work, as the guidance recommends.
- Credential or MFA reset: second-person approval and confirmation through a previously registered channel
- Contact-detail change: step-up verification and a hold on further sensitive changes
- Authorized-user addition: approval from the account holder through a trusted channel
- Verification override: always escalated, never completed by the requesting agent alone
Where runtime policy enforcement applies to AI-assisted tooling
AI assistants in agent workflows summarize calls, suggest next steps and call tools such as account lookup, contact update or reset. The same high-risk actions can now be triggered through an assistant. Content that originates outside the trusted system prompt, including anything the caller says or types, should be treated as untrusted. OWASP guidance on agentic AI also advises logging every agent action with its originating input.
Policy enforced in the execution path decides each interaction or tool call before it completes. It can allow the call, redact sensitive fields from what the assistant sees or returns, block the action, or escalate it for human approval. Enforcement should sit outside the assistant, because a control the assistant can ignore is not a control. It also depends on an inventory of the tools available to assistants, with their permissions and usage policies, since policy cannot apply to a tool nobody has listed.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. Enforced at runtime, not reviewed after the fact, a reset or contact-change tool call can be held for human approval. The conditions that trigger that hold, such as verification state or a fraud score, must come from your identity and fraud systems.
Evidence an examiner can review, and what enforcement does not replace
Decide in advance what each decision record must contain: the originating input, the tool and parameters, the human agent identity, the AI agent identity, the policy version, the outcome, and for escalations the approver and time. NCCoE material describes linking specific user identities to AI agents so that accountability is preserved, which is what lets an examiner attribute an AI-assisted action to a person.
The OCC supervisory framework, which uses permissive language, links control intensity to the consequences of agent failure or misuse. It describes documenting and testing detection, containment, suspension and escalation arrangements, and providing evidence that controls keep operating. NYDFS guidance of October 2024 names social engineering among AI-related cybersecurity risks and takes a risk-based approach. Decision records from enforcement can support that kind of assurance. Trussed produces audit-ready evidence from enforcement, mapped to NYDFS and NIST, with configurable retention and self-managed deployment in the customer cloud. Trussed does not have financial-services deployments, and it is not a source of fraud signals.
Enforcement does not replace fraud detection, which identifies the pattern. It does not replace identity verification, which establishes who is calling. It does not replace case management, which holds the investigation. Its role is to constrain the action at the point it would execute and to leave a record of what was allowed, changed, blocked or escalated.
From manipulated call to reviewable evidence
Caller verification
Failed, weak or overridden identity checks at the start of the call.
Agent behavior
Session activity that departs from the agent's normal service pattern.
Account change pattern
Contact, credential and access changes in a risky order.
Enforcement and evidence
Allow, redact, block or escalate at the point of action, with a decision record.
Constrain and evidence high-risk agent actions
See how policy enforced in the execution path handles AI-assisted tool calls and produces audit-ready evidence for examiner review.
See How Evidence Is Produced