Why help desk social engineering beats good authentication
Every control discussed in passkeys and phishing-resistant authentication works on the sign-in. None of them works on the enrolment. If an attacker can get a new credential issued to themselves, the strength of the credential they bypassed is irrelevant, because they are now holding a legitimate one.
This is not a theoretical concern. A joint CISA and FBI advisory describes an actor set that "targets large companies and their contracted information technology (IT) help desks", and that uses "social engineering techniques to convince IT help desk personnel to reset passwords and/or MFA tokens". Its July 2025 update adds the mechanism in one sentence, describing actors who "Posed as employees to convince IT and/or helpdesk staff to provide sensitive information, reset the employee's password, and transfer the employee's MFA to a device they control on separate devices." The same advisory records the reverse direction, actors posing as IT or help desk staff to employees over phone calls or text messages to obtain a one-time code, and then registering "their own MFA tokens". It also records that the same actors convinced mobile carriers to transfer a target's phone number to a card they controlled, which defeats any verification that relies on calling the number on file.
What the documented cases actually look like
Two cases are worth reading because the mechanism is recorded call by call rather than summarised.
The first is a civil claim, so treat everything in it as allegation. A consumer goods manufacturer is suing the provider that ran its outsourced help desk over an August 2023 intrusion. As reported from the complaint, a caller asked for a reset of an employee's identity provider password; when asked to connect to the VPN, said they did not have that password either; and the agent reset both without further questioning. The caller then said their multi-factor authentication was not working, and that was reset too. A second multi-factor reset was granted the same day. A third call asked for the identity provider password to be reset again. The caller also asked for the registered phone number to be changed so codes would arrive at a number they controlled, and that was granted. The central allegation is one sentence: "At no point during any of the calls did the Agent verify that the caller was in fact Employee 1."
Two details in that complaint matter more than the headline. First, written procedures existed. The company's own guidelines required the use of its self-service reset tool, or, where that was unavailable, asking the caller for the manager's name and the account username. So the failure was not an absence of policy. It was policy not followed under pressure, which is a different problem with a different fix. Second, the alleged damages are 380 million US dollars, and the complaint states the intrusion was detected after three hours. Three hours of detection latency did not help, because the attacker already held valid credentials.
The provider disputes the claim. Its position, quoted in coverage, is that it "was hired for a narrow scope of help desk services which Cognizant reasonably performed" and that it "did not manage cybersecurity" for the customer. Whatever a court decides about liability, the mechanism described is the one to defend against.
The second case is a company's own published post-mortem from September 2023, which is more useful because it is self-reported. Several employees received text messages claiming a member of IT was reaching out about an account issue, linking to a page disguised as the company's internal identity portal. The timing matched a recent migration to a new identity provider, which made the story plausible. One employee signed in, and the fake portal collected a multi-factor code. The attacker then phoned that employee, claimed to be from the IT team, and used a synthesised copy of a colleague's voice. The employee grew suspicious during the call and still handed over one more code. That single code, in the company's own words, "allowed the attacker to add their own personal device to the employee's Okta account."
There was a compounding factor the company is blunt about. A cloud sync feature in the authenticator application meant that access to one cloud account gave access to every code held in it, so what administrators believed was multi-factor authentication had, in their words, "silently (to administrators) become single-factor-authentication". This is the same category of problem described in MFA bypass techniques, arriving through configuration rather than through an attack.
What the standards actually require
NIST's digital identity guidance is clearer on this than most people expect. On recovery it requires that the provider "repeat the necessary steps of identity proofing consistent with the level of initial identity proofing" and "confirm that the claimant's identity is consistent with the previously established account". Where an account was never identity proofed in the first place, which covers most ordinary business accounts, reproofing is not available, so recovery must instead "require the successful use of a saved recovery code, issued recovery code, or recovery contact".
Three further requirements are worth knowing. Setting or changing a pre-registered telephone number is treated as the binding of a new authenticator, not as an administrative detail, which means it carries the same authentication requirements as adding a factor. Records of significant authenticator lifecycle events must contain the date and time, and should include the source of the binding such as an address or device identifier. And from the non-normative material, the principle that underlies all of it: "it is essential that one authentication factor cannot be leveraged to obtain an authenticator of a different factor."
Note the recurring answer to recovery in both this guidance and the vendor documentation. Nobody recommends making recovery stronger. They recommend reducing how often you need it, by having people enrol a second authenticator in advance.
There is a gap worth naming, because it shapes what you can look up. The CISA and FBI advisory describes this attack across three versions, most recently updated in July 2025, and its mitigations answer with phishing-resistant authentication using FIDO or WebAuthn. That recommendation is sound and it does not address this attack, because an attacker who persuades a help desk to enrol a factor for them ends up holding a phishing-resistant credential of their own. The mitigations say nothing about verifying who is calling, handling a reset request, or re-enrolling a second factor. We checked the current version on 16 November 2026. So there is no authoritative procedure to copy, and the recommendations below are assembled from vendor settings, standards requirements and reasoning, each labelled as such.
The settings that are documented, and the ones that are only sensible
Start with the documented ones, because they are free and specific.
Require two methods rather than one for self-service reset. Microsoft states that the number of methods required "can be set to either one or two". Set it to two. Be aware of two defaults: administrator accounts are enabled for self-service reset by default under a two-gate policy that "requires two pieces of authentication data" and "prohibits security questions", and that administrator policy cannot be changed. Meanwhile the weaker one-gate policy, requiring a single piece of authentication data, applies where no custom domain is configured on the tenant, which is a common state for small deployments.
Switch on administrator notification. Microsoft's setting is named "Notify all admins when other admins reset their passwords", and it emails global administrators when another administrator changes their password by self-service. If you build one alert from this article, build that.
Turn off self-service reset for administrators if you can run without it. Microsoft documents this as a tenant authorization policy property, and notes that changes can take up to 60 minutes to take effect. In Google Workspace the equivalent is the super administrator self-recovery setting: with it off, a super administrator who forgets their password "must contact another super admin or Google Support". Google recommends enabling self recovery only where there are multiple super administrators to help each other, and advises removing a departing super administrator's recovery information immediately.
Use a time-limited pass for genuine lockouts rather than a reset. Microsoft's Temporary Access Pass is a time-limited passcode, configurable for single or multiple use, intended for exactly this situation. Read its limitations before relying on it: each user can have only one, it does not replace the user's password, and a session established with it is not invalidated when it expires, so someone can retain access longer than the pass lifetime unless sign-in frequency is configured.
Remove the old method after recovery, and review for methods you do not recognise. Microsoft's own guidance says users who update their authentication methods after losing a credential or device "should make sure they remove the old authentication methods", and its compromised account guidance tells administrators to identify and remove any suspicious devices or unrecognised methods an attacker added. For Google Workspace, the compromised account steps cover suspending the user, which Google says resets sign-in cookies and OAuth tokens, then revoking OAuth tokens, removing application passwords the user created, and reviewing recovery addresses. What survives a credential change is covered in what survives a password reset.
Now the ones that are reasoning. Calling back on the number already on file, refusing to act on inbound contact at all, imposing a short waiting period on credential changes, and requiring confirmation from a named manager are all sensible and all widely recommended, and we could not find authoritative published guidance behind any of them. We are presenting them as reasoning, not as a framework, because this site does not dress inference up as citation. Each also has a failure mode. Call-back fails against a carrier transfer of the number, which the CISA advisory records as in use. Manager confirmation fails when the manager is the person being impersonated. A waiting period fails when the request is genuine and urgent, which is how exceptions get made and procedures get abandoned.
One link to your own exposure data, labelled honestly as reasoning: an attacker who already holds an employee's leaked personal details, work email format and job title can answer low-entropy verification questions convincingly. No source we read documents attackers sourcing help desk answers from breach data, so we are not going to claim it. What is documented is that the verification questions in the case above were exactly that kind of question, and that Microsoft prohibits security questions for privileged resets. Draw your own conclusion about what your help desk currently asks.