How Google Stops Account Hijacking With the Right Challenge — and How to Do It in Your Business
Most hijacked accounts are not hacked in any clever way. Someone has the right password, from a breach or a phishing page, and types it in. To your login form, that looks like the real user.
Google faces this at a scale few companies do. Its own security blog describes protecting users from hundreds of thousands of hijacking attempts a day. Blocking every unusual sign-in is not an option: people travel, buy new phones and change networks all the time. So Google asked a narrower question: when a sign-in looks risky, which extra step actually stops the attacker, and what does it cost the real person?
What Google measured
The answer was published at The Web Conference (WWW) 2019 by Periwinkle Doerfler and Damon McCoy (New York University) with Maija Marincenko, Juri Ranieri, Yu Jiang, Angelika Moscicki and Kurt Thomas (Google): Evaluating Login Challenges as a Defense Against Account Takeover (DOI). Google summarized it on its Security Blog in May 2019: New research: How effective is basic account hygiene at preventing hijacking.
The method, plainly:
- Score every sign-in. Google’s risk engine judges IP reputation, device reputation and automation patterns. Most sign-ins pass silently.
- Challenge only when it looks wrong. The paper calls this risk-aware authentication: a challenge appears “in response to a perceived threat”, not on every login. In the phishing sample, every challenge was triggered by a sign-in from a country the account had never used.
- Pick the challenge the account can support. The study compared 14 challenges in four families: device-based (a prompt on a phone already signed in, an authenticator code, a security key, a code to the registered phone), delegation-based (a code to a backup email), knowledge-based (last sign-in location, a security question, the phone number on file) and resource-based (a code to any phone or email the person supplies).
- Measure both sides. For attackers, the share of challenges they failed. For real users, the share who passed, skipped or gave up.
The data covered more than 350,000 real hijacking attempts between February and May 2018: 300,000 from automated bots, 50,867 from phishers and 453 from targeted attackers. For usability, the team sampled 1.2 million challenges solved by legitimate users.
The numbers they published
From the paper’s abstract and results:
| Challenge | Automated bots stopped | Phishing-rooted attempts stopped |
|---|---|---|
| Device-based challenges (as a family) | 100% | Over 94% |
| Prompt on a phone already signed in | 100% | 99% |
| Security key | 100% | 100% (small sample) |
| Code to a registered backup email | 100% | 92% |
| Knowledge-based challenges (as a family) | As low as 73% | As low as 10% |
| “Where did you last sign in?” | 100% | 10% |
Across all sign-ins, the paper reports that Google’s risk engine blocked over 99.99% of automated hijacking attempts and 92% of attacks rooted in phishing.
Two details explain the gap. Bots fail device challenges because they have no access to the real person’s device. And a phishing page can simply ask the victim where they last signed in, then type the answer into Google. A prompt on the phone is harder to fool, because it shows the account holder which device and location is trying to get in. Phishers got past 1% of those prompts.
The cost is real too. The abstract says friction led 52% of legitimate users to fail a challenged sign-in, though 97% eventually got into their account in a short period. The blog adds that in one experiment, 38% of users did not have their phone with them when challenged, and 34% could not recall their secondary email address. That is why Google challenges only risky sign-ins, and why the blog notes that a user who fails a challenge can sign in again from a device they have used before.
What transfers to your business, and what does not
Be honest about the gap. Google has billions of accounts, its own threat intelligence feeds, and a phone in many users’ pockets that is already signed in to Google, which is what makes the on-device prompt possible. Most small SaaS products, stores and gaming sites have none of that. You probably have no app on your users’ phones to push a prompt to.
What does not transfer:
- Push prompts to a signed-in phone, unless you already have a mobile app people stay logged into.
- The scale of the measurement. You will not have 350,000 attacks to tune against, so borrow the conclusions.
What does transfer:
- Challenge on risk, not on every login. A familiar device on a familiar network should see nothing. The friction numbers above are the reason.
- Knowing the device is the strongest passive signal you have. Google leans on familiar devices and networks to decide when not to challenge. A device identifier gives you the same knowledge without asking the user anything.
- Skip knowledge questions. “Where did you last sign in?” stopped 10% of phishing-rooted attempts. Security questions have the same weakness: whatever the victim knows, a phishing page can ask for.
- Prefer a step an attacker cannot relay. Security keys stopped every attack in the sample because the key checks which site is asking. Passkeys, now built into phones and browsers, work on the same principle. The 2019 study did not test passkeys, so treat that as reasoning, not a published result.
- Keep a way back in. Some real users will fail. Plan the fallback before you turn anything on.
How to do it in your business
Prynt Control turns this into a login decision you can apply on day one. Prynt never blocks on its own: it returns a decision and the step it recommends, and your app decides what to show.
- Recognize the device at login. Prynt identifies the browser or phone and tells you whether this account has used that device before. A known device on a usual network gets no step, which is what most of your real users should see.
- Use the login decision. It is one of Control’s ready-made use cases, with three outcomes: allow, step up or block. Its inputs follow the same idea as Google’s engine: is this the account’s first login from this device, did the device change country within the last hour, has the account been used from many devices in a day, are there many attempts from one device in a few minutes, are there automation signals.
- Climb the friction ladder one rung at a time. Instead of a yes or no, the answer names the step: none, captcha, verify email, card or passkey, block. Automation signals and bursts of attempts point to a captcha, since bots fail anything that needs a real person. A new device for a known user points to an email check. A country change within an hour, or an account used from many devices in a day, points to card or passkey: a passkey if the user has one. Block is kept for direct evidence, such as a device or account you labeled fraud or one on your deny list. When the user passes the step, report it with
POST /v1/control/steps, and that device is not asked again at login for 30 days unless something stronger shows up. There are no security questions on the ladder, and no codes sent by text message. - Watch before you act. For the first 14 days, mirror mode only observes, then shows case by case what the login decision would have challenged or stopped. Check how many were your real customers before you turn anything on; Google’s 52% figure is the warning. Later rule changes can run in evaluate mode first.
Device signals have limits, and we publish ours on the accuracy page. In our lab, Prynt re-identified the same device 100% of the time, detected 89% of stealth automation tools (238 of 268) and flagged 0 of 240 real humans as bots. Lab results are not field results, so start in mirror mode.
The honest part
Google’s challenges did not stop everything. Targeted attackers still got through some of them, and every challenge cost some real users their sign-in. What the research showed is that the choice of step matters more than the number of steps: a challenge tied to something the attacker does not hold beats a question the victim can be tricked into answering.
You don’t need Google’s scale to use that. You need to know the device, challenge only when something looks wrong, and choose a step a phishing page cannot pass along.
Prynt is not affiliated with Google or the paper’s authors; this post describes their published work.
Try it free
Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.