Anchor

## Inspiration

Scammers don't need to break into accounts anymore. They break into people.

The HuskyHack starter kit introduced us to **Linda**, a retired teacher who gets a panicked call from her "grandson," then a second call from his "attorney." She's told to keep it secret and tries to wire **$8,500**, her first wire in over six years. We also met **Sarah**, who clicks a text that looks like it came from Sound, enters her password and one-time code on a fake site, and loses control of her account within minutes.

Two things stood out to us:

1. **Generic warnings don't work.** A member under pressure clicks past "Beware of scams" in seconds.
2. **The fraudster often controls the confirmation channel.** In Sarah's case, every code Sound sends to the "new" phone goes straight to the attacker.

We wanted to build something that steps in at the exact moment money moves, respects the member's right to decide, and stays invisible to members whose payments are normal.

## What it does

**Anchor** has two modules that share one risk engine and one employee console.

### Module 1: Pause & Reflect (before Send)

When a payment breaks the member's *own* pattern, Anchor pauses it before it leaves:

- **Risk tiers based on the member's history.** A routine large payment sees no friction. Linda's first wire in six years to a new recipient triggers the full flow.
- **A critical questionnaire** that exposes the scam script: *Were you asked to keep this secret? Were you told what to say if Sound asks? Have you called your grandson on a number you already had?*
- **A "Check my risk score" button** that calls our backend and returns an explainable 0–100 score with plain-language reasons.
- **A trusted-contact alert** sent before Send to family or beneficiaries the member chose to enroll. Contacts are **informed, never given control**.
- **An employee callback** inside the 4-hour hold, with a suggested non-judgmental script.

The member always makes the final decision on their own payment.

### Module 2: Takeover Lockdown (after a takeover is confirmed)

Once Sound confirms an account takeover, Anchor runs a five-step lockdown:

1. **Delink** digital access from every money action (sending, adding recipients, opening accounts).
2. **Contain** by cancelling pending payments, revoking sessions and reverting contact changes.
3. **Re-verify** identity through channels the fraudster doesn't control: a branch visit with photo ID or a callback to the *pre-change* phone number.
4. **Migrate** the member's funds to a brand-new account, redirecting payroll and autopays.
5. **Demolish** the compromised account so it can never move money again, then request wire recalls and ACH returns.

## How we built it

- **Member app and employee console:** Next.js and TypeScript, covering the transfer screen, questionnaire, risk result, trusted-contact alert and migration status screen.
- **Backend:** a .NET API exposing `POST /v1/risk/score` and `POST /v1/alerts/trusted-contacts`, with PostgreSQL for cases and the audit log.
- **AI:** Azure OpenAI classifies the member's free-text answers into known scam typologies.
- **Data:** the HuskyHack synthetic events file and Member Key, used to replay each journey and to test triggers against the legitimate look-alike members.

The risk score combines three signals: a pattern score $P$ from the member's transaction history, an answer score $A$ from the questionnaire, and a narrative score $N$ from the AI model:

$$
\text{Score} = 100 \cdot \sigma\left(w_P P + w_A A + w_N N\right)
$$

The answer score weights each red-flag answer $a_i \in \{0, 1\}$ and adds a penalty for contradictions $c$ between answers:

$$
A = \sum_{i=1}^{n} w_i \, a_i + \lambda \, c
$$

Scores of 70 or above map to the high tier, which triggers the trusted-contact alert and an employee callback. Every score comes with reason codes such as `SECRECY_REQUEST` and `FIRST_WIRE_6Y`, so members and employees can see *why*.

## Challenges we ran into

- **Friction versus safety.** We had to stop scams while keeping legitimate payments easy. Scoring against each member's own history means friction appears only when a payment truly breaks their pattern.
- **Autonomy.** For a member's own payment, Sound can pause and advise but can't block. Every screen keeps the final say with the member, and "Send anyway" leads to a human conversation instead of a dead end.
- **The compromised confirmation channel.** In a takeover, codes sent to the new phone or email reach the fraudster. We built verification around the contact details from *before* the change, plus in-person verification.
- **Trusted contacts can be part of the problem.** Contacts receive only a minimal notice, and the alert is suppressed for employee review if a contact is linked to the recipient.
- **Staying within realistic data.** We left out device intelligence, voice analysis and access to members' messages, and built only on data a credit union already has.

## Accomplishments that we're proud of

- **One platform covering two very different scams:** an authorized-payment scam where the member must keep control, and an account takeover where Sound must act fast.
- **Every role is designed in:** the member, the employee on the case and the trusted contact each have their own screen.
- **Explainable AI:** every score ships with reason codes a member or employee can read, never a black-box number.
- **A fast legitimate path:** a look-alike member wiring a title company for a home closing gets through with three questions instead of a hold.
- **A complete recovery story:** if prevention fails, Anchor moves straight into recalls, account migration and follow-up check-ins.

## What we learned

- **A risk score isn't a solution.** The value is in the *next best action*: the right question, the right person to alert, the right moment to pause.
- **Good scam prevention is mostly UX.** The questions that work are specific, non-judgmental and written for someone who is scared and in a hurry.
- **Speed matters most after prevention fails.** Wire recalls and ACH returns get less likely every hour, so recovery has to start the moment a takeover is confirmed.
- **Designing for false positives is designing for trust.** The synthetic data has far more scam victims than real life, so every trigger needed a fast, respectful way through for legitimate members.

## What's next for Anchor

- **A 90-day pilot with Sound:** 30 days in shadow mode to measure false positives, 30 days with Pause & Reflect live on wires and external transfers, and 30 days with Takeover Lockdown live alongside the fraud team.
- **Branch and Contact Center versions** of the questionnaire, so tellers can run the same check before large cash withdrawals.
- **More scam typologies,** including romance and government-impersonation scams, on the same risk engine.
- **Multi-credit-union rollout** through a shared credit union service organization, so smaller credit unions can offer the same protection.