The helpdesk is the new perimeter

Nobody attacks the strong path when a weaker one reaches the same account. Every hardening step of the last decade pushed effort somewhere else, and the destination it kept arriving at was the set of flows that exist to restore access after authentication has already failed. Security questions went first, once breaches and social media turned private answers into lookups. One-time codes moved the problem onto a phone number, and numbers turn out to be transferable. Push approval moved it onto a tap and met fatigue. The current rung is the session itself, relayed whole by adversary-in-the-middle kits, which is the category the Microsoft Digital Defense Report measured growing 146 percent year over year in 2024.

Under every rung is the same fallback. When someone cannot get in, a person on a support line decides whether to let them, working from a script, whatever the account record shows, and an instinct to be helpful. That fallback is now attacked deliberately rather than opportunistically. In 2023, attackers impersonated an employee to the MGM Resorts IT helpdesk by telephone, and the company reported an about $100M impact. No cryptography failed anywhere in that chain.

Then came the escalation that recognition cannot survive. In February 2024, a finance employee at the engineering firm Arup joined a video call with what appeared to be colleagues, including the chief financial officer, and authorized transfers of about US$25M on their instruction. Every face and voice on that call was synthetic. Once a caller can present the right face and the right voice on demand, recognition stops being weak evidence and stops being evidence.

Since the failure is built into the shape of the flow, the remedy on offer is not a sharper script or a longer interview. Making an agent feel more confident is the wrong project. The position taken here is that agent confidence should stop functioning as a control at all, replaced by a question no rehearsed story can answer.

That question is one the caller's own phone can answer while the call is still in progress. SenseCrypt exposes it as a backchannel verification built on FAPI-CIBA / CIBA, the OpenID standard written for flows that begin on one channel and finish on another. The check consults the person rather than the record, so nothing is read out of the account notes and nothing is asked of the caller.

The paper is careful about why the direction of the check matters. Anything verified inside the call is verified over a channel the caller controls, which is exactly why synthetic audio defeats voice matching where voice matching is deployed. A check that leaves the call, completes on hardware the caller cannot influence, and returns as a signature does not inherit the compromise of the conversation.

1. Each control on the escalation ladder and the attack that retired it, ending at the fallback everything else rests on
2. What the FAPI-CIBA / CIBA exchange looks like from the agent's side, from one click to a result on screen, and how long it takes
3. Why the description of the action being approved travels only inside the signed payload, and what that prevents
4. The route for a caller who has never set up the app, and why the mailbox in that route is a bootstrap rather than a substitute
5. What deprovisioning does to a departing employee's ability to pass the check, and at what moment it binds
6. Three things a fraud team should expect to change, among them how a disputed action reads once an investigation opens

Fraud teams, the people who own a contact center, and whoever ended up holding recovery once the authentication program was signed off. The premise is uncomfortable and worth sitting with: the strongest control in the estate is not the one an attacker will choose to meet.

It is also written for the people who have to explain an incident afterward. Every verification, passed or failed, lands in a read-only person-level activity stream that investigators can export to CSV, which turns a reset from a line in a ticketing system into a record of which identity was checked, with what outcome, and when.

Seven pages and two diagrams, with incident figures taken from what the companies themselves disclosed and from reporting at the time rather than from vendor modeling. The section on what changes for the fraud team is the one to take into an operations review.

The gated download sits on sensecrypt.com, and a copy of the PDF is available on request. The more useful next step is a live walkthrough: trigger a check against a test identity, watch it complete on a phone in the room, and see what reaches the agent's screen when it fails.