Vaultish employee approves 47th MFA prompt after her manager's voice asks nicely
The manager was asleep. The voice was not, and it had already called the help desk.
By Marguerite Okonkwo-Hale, Senior Breach Correspondent4 min read
At 2:14 a.m. on a Tuesday, a payroll specialist at Vaultish, a fintech company whose app promises to "lock down your money," received a push notification asking her to approve a sign-in. She declined. She received another. She declined that one too. She declined the next forty-four.
"Around number thirty I started to wonder if it was a test," said the employee, who asked to be identified only as "the person in the incident report." "Around number forty-six my manager called to tell me it was neither."
The caller sounded exactly like Priyanka Desrosiers, the payroll manager the employee has reported to for three years. It explained that IT was migrating the authentication system overnight, that the prompts were a "sync loop," and that the fastest way to clear it was to approve the next one and read back the six-digit code so the ticket could be closed. The employee approved prompt forty-seven and read the code. The voice thanked her, told her to get some sleep, and hung up.
Desrosiers was asleep. Her voice, investigators later established, had been assembled from four years of recorded all-hands meetings and a training video. The call was placed by an automated voice agent operated by the attacker, which had been given the employee's name, her manager's name, and a script.
The agent had been busy. Fourteen minutes earlier it had called Vaultish's outsourced IT help desk, identified itself as Desrosiers, said it was locked out while traveling, and asked for the recovery email on her account to be updated. The help desk asked for her employee ID and start date. The voice supplied both, read off a professional networking profile and a welcome post on the company blog. The help desk changed the recovery email and noted: "Caller sounded a bit robotic but was very polite."
Vaultish's internal post-incident review, noted that the attacker had obtained the employee's password from a credential-stuffing list three weeks earlier, and that it was "Vaultish2022!"
The review further noted that the company's authentication provider offered number matching, which would have required the attacker to see a code displayed on the real login screen, and a setting to limit push notifications to three per ten minutes. Both had been disabled in 2024 "because they were annoying during onboarding."
"We have reminded the employee never to approve a prompt she did not initiate. We have reminded the help desk that callers may not be who they sound like. We have not yet reminded ourselves that the attacker made 48 phone calls that night and we picked up every one." — Vaultish incident review, page 4
Vaultish has since rolled out a security awareness module titled "Is That Really Your Manager?", which is 35 minutes long and narrated by a synthetic voice.
The part that is not a joke
What actually happened
This is MFA fatigue, also called push bombing, with the second stage automated. Push-based MFA asks a tired human a yes-or-no question with no context, over and over, until the answer changes. The attacker already holds a valid password from a breach dump or an infostealer; the second factor is the only thing in the way.
The phone call used to be the expensive part. A voice agent built on a cloned sample does that at scale, in the target's own manager's voice. The same agent calls the help desk, answers the verification questions from public data, and changes the recovery path so the account stays lost after the password is reset.
Nothing in that chain needed a new vulnerability. It needed push MFA without number matching, a help desk that trusts a familiar voice, and no limit on prompts.
What to do about it
For organizations:
- Turn on number matching, so the user must type a code shown on the real login screen into the authenticator. A voice agent reading a script cannot see that screen.
- Rate-limit push prompts and alert the security team when a user denies several in a row. Forty-seven denials is a signal, not noise.
- Move the highest-risk roles to phishing-resistant MFA: FIDO2 security keys or passkeys. There is no prompt to approve and no code to read aloud.
- Rewrite help-desk verification on the assumption that any caller may be synthetic. A familiar voice is not a factor. Require a registered device, a callback to a number on file, or a manager's confirmation on a separate channel for every reset or recovery change.
- Tell staff, in writing and often, that IT and managers will never call to ask them to approve a prompt or read out a code. A call that sounds like someone is not evidence it is them.
For individuals:
- Never approve a prompt you did not just trigger. If prompts keep coming, your password is compromised. Change it tonight, everywhere you used it.
- Hang up on anyone who calls about an MFA prompt, including your boss, then call back on a number you already have.
- Stop putting the company name in your password. The attackers know where you work, and now they know what your manager sounds like.
About the writer
Senior Breach Correspondent
Marguerite has read more breach notification letters than any living person and has started to dream in the passive voice. She covers incidents, disclosures, and the gap between the two.