On Sunday night, I received an email from Coinbase notifying me that someone was attempting to log into my account. A few seconds later, two separate 2FA codes arrived on my phone (the short one-time passcodes apps send to verify it’s really you), neither of which I had requested.

Less than an hour earlier, I had gotten off the phone with someone I believed was a Google support representative. The call came from Google’s actual outbound number. The caller knew enough about Google’s recovery process to sound completely legitimate, and by the end of the conversation I was convinced I had just helped secure my account from an attacker. Instead, I had handed control of it to one.

The Phone Call

This is called social engineering: an attacker manipulates you psychologically instead of hacking anything technically. There was no malware and no stolen password, only a convincing script and enough pressure to make me act before I thought.

The call appeared to come directly from Google. My phone displayed their real outbound number, which is a spoofing trick: the call shows a number it isn’t really coming from. An automated voice told me someone was trying to change the recovery email on my account.

A recovery email is a backup address Google uses to verify your identity if you get locked out. Changing it lets an attacker take permanent control.

I was transferred to a live “representative” who told me a hacker was actively trying to compromise my account and that he needed my help to stop it. The script was effective because it made me feel like I was already a victim, when the attack was the phone call itself.

Over eighteen minutes, he walked me through what looked like legitimate Google security steps, coaching me to read verification codes aloud that were actually granting him access. Every email I received looked real because it was real. Google was sending them. I just didn’t understand what they were actually authorizing.

By the end of the call, he had changed my password and set up a new recovery path. He had also created several ways back into the account, built to survive even if I figured out what had happened.

The Coinbase Alert

About forty-five minutes later, the plan started to unravel. With access to my inbox, the attacker found Coinbase and attempted to trigger the account recovery process there too. That set off security alerts and 2FA codes sent to my physical phone, which he didn’t have. Coinbase flagged the suspicious activity, automatically froze the account, and the assets stayed safe.

But the Coinbase alert emails were nowhere in my inbox. That didn’t add up.

The Audit

This is where the AI agent became critical. I was focused on the obvious things: changing my password, checking which devices were logged in, reviewing login history. The AI agent started asking a completely different set of questions.

What passkeys exist? What app passwords exist? What recovery relationships exist? What email filters exist?

It was looking for evidence that the attacker planned to come back.

It directed me into a section of Gmail settings I had never opened: Filters and Blocked Addresses. Email filters are rules that sort, label or hide incoming messages automatically. They’re useful for organizing your inbox, and apparently useful for hiding evidence of a crime.

That’s where we found it. The attacker had created rules targeting financial institutions: any email containing words like “Coinbase” or “Wells Fargo” was marked as read and archived before it ever reached my inbox.

The attacker’s filters were hiding the alerts as they arrived.

The attacker was building a blind spot: a quiet blackout where security alerts, fraud warnings, and password reset emails would disappear while everything looked normal from my end. We deleted every malicious rule, and the blackout ended immediately.

Finding the Backdoors

The filters weren’t the only thing left behind. The agent walked me through the other ways an attacker can keep access to an account after a password change.

We found an unauthorized passkey (a modern login credential tied to a physical device, designed to be more secure than a password). Left in place, it would have been a working key to the account regardless of what else was changed. We removed it.

We found an app password (a special access token generated so that external apps can connect to Gmail without needing the main password each time). The attacker had created one during the call. We removed that too.

Then we invalidated all active sessions globally, signing out every device at once, and removed the recovery relationship established during the phone call.

Every way back in that we could find was closed.

The False Alarm

The next day I spotted a generic Android device in Google’s device dashboard that appeared to show a recent sync, months after being authorized. Given everything that had happened, it looked like a new threat.

The AI agent reconstructed the timeline and caught what I had missed. Its reading was that the behavior matched Google’s own cleanup: after a password reset, previously authorized devices get signed out. What I was seeing was an old household device being kicked out. The cleanup had worked.

Locking It Down

With the account secured, we looked at making it harder to compromise in the first place. We turned off two-factor codes by text message, which an attacker who takes over your phone number can intercept, and switched to stronger verification.

We also added a YubiKey, a small physical USB key that acts as a second form of ID. Signing in now takes the key itself, so someone with only my password can’t log in.

One detail I noticed: Google made me wait before the new hardware key could be used for account recovery. It felt like friction at first, but the logic is sound. The way I read it, if an attacker adds their own key, you have a window to catch and remove it before it works.

What I Learned

The attacker got into my Gmail and nothing else. Coinbase, my financial accounts and my assets were all untouched.

I didn’t handle everything well under pressure. The situation was moving faster than I could follow, and I was making decisions without fully understanding what was happening.

What stopped the damage from spreading was that the investigation was more systematic than the attack. The AI agent approached it like an incident response analyst, someone trained to document, audit and close off an active breach, while I was the panicked account owner checking the obvious things. Its value was knowing where to look after the password had already been changed.

An Agent You Can Ask in the Moment

What protected me was an AI agent that could think through a crisis methodically when I couldn’t.

In Archie, that’s an agent you can ask what to check, in the moment, and it answers the way any AI chat does. Check what it tells you, because like any AI, it is sometimes wrong.