Layered Authentication: stop Identity Fraud without taxing every User

1. Introduction: boring Half of Identiverse#

If you read any recap of Identiverse 2026, it was about agents. Every keynote, every booth and most of the interview stage. We covered that side in blended identity and runtime authorization and in shadow AI and agent debt.

But there was a quieter thread that matters more to anyone running consumer logins. It is the same thread we were in Las Vegas for: our own session with VicRoads walked through taking passkeys live for more than 5 million users at a 70-90% activation rate. The fraud track was its natural counterpart.

The framing that cuts against the AI hype came early. Asked how much AI is changing fraud, the answer split in two: AI is enabling genuinely new attack types like deepfakes, but the bigger effect is as a force multiplier on schemes that have existed for a decade.

That distinction matters, because it decides where you spend. If it is mostly the same attacks executed faster and at higher volume, then the answer is the unfashionable one: get the fundamentals of authentication and risk assessment right, and make them work at scale.

2. There is no silver Bullet, there is a Ladder#

The central recommendation was layered authentication, and the fraud sessions were specific about what that means operationally.

The sequence: assess passively first, using signals the user never sees. Device reputation, behaviour and traditional identity signals such as name, address and date of birth. Then, depending on the risk of that specific session and the risk of the transaction being attempted, decide whether to add friction and how much.

The anti-pattern is the one most organisations actually run: a single workflow applied uniformly to every customer. Nobody should have to answer knowledge-based questions or receive a one-time password every time they check an insurance policy.

The internal business argument is worth stealing. Fraud is a small percentage of transactions if you are doing reasonably well, so a control that taxes 100% of users to catch it is a bad trade. Framing it as balancing fraud against customer experience is what gets it approved.

This is the same logic behind passkeys as a default path rather than a step-up mechanism: the low-friction option should be the secure one, so that added friction can be reserved for the sessions that actually warrant it. The following diagram shows how a single session is routed through that ladder.

3. Fraud is a Lifecycle Problem, not a Login Problem#

The second theme was that assessing identity once, at account creation or at login, is not enough. The analogy used was a house with a front door, a back door and windows, where most fraud programs secure one of them.

The insurance example walks the full arc. A consumer arrives on the website for a quote, and that is the first point where device risk and identity need to be assessed, before any policy exists. Months or years later the same identity comes back to file a claim. If the verification happened properly at quote time, the claim can be matched against it. If it did not, the claim is the first real check, and by then money is moving.

The timeline below lays that arc out and marks where the silos break it.

That also corrects a common assumption about why insurance and healthcare get attacked. The old thinking was that they are less exposed than financial services because there is less direct money to steal. The actual prize is non-public data:

The organisational blocker is silos. Claims has a claims problem, underwriting has a bot problem on the quoting platform and consumer experience has a login problem. All three are the same person.

CIAM teams know that shape. Login data lives with one team, fraud signals with another and support tickets with a third, so nobody sees the whole path a user took.

4. Weakest Handoff, not the weakest Door#

The session that most directly confirms the argument in section 3 stated it more sharply than we would have dared:

The lifecycle map is the one to steal. Five stages, each with its own control and its own characteristic failure:

The conclusion is the sentence every CIAM team should have on a wall:

Note where a passkey sits in that table. It hardens exactly one row, and it can be bound by the wrong person in the row above it. Enrollment binding is the step where a strong credential gets issued to a fraudster, after which every subsequent login is cryptographically perfect and completely wrong.

4.1 Numbers behind it#

The figures the session was anchored to, sourced from TransUnion, Thomson Reuters, LexisNexis Risk, the Financial Times and Signicat:

The last two rows are the business case for layered authentication stated from the revenue side rather than the security side. Two thirds of applicants abandon onboarding, and nearly half of the people who hit friction take their money elsewhere. A control that adds friction uniformly is not neutral, it is expensive.

The characterisation of the modern attacker explains why volume has changed:

5. Deepfakes: what actually changed#

On the genuinely new side, one control pattern has settled for document-based verification: scan the document, check it is real and unmanipulated, take a selfie with a liveness check to confirm a real person is present in real time rather than a synthetic video, then match the two.

The external benchmarks put scale behind this:

The important nuance for authentication teams: deepfakes attack identity verification, the process of proving who a new or unknown person is, far more than they attack authentication, the process of proving a returning user is the same person as before. A passkey is scoped to the domain it was created for and proves possession of a private key, so it does not care what a face looks like on camera. Synced passkeys, which are backed up and available across a user's devices, behave the same way here as device-bound ones. That does not make deepfakes irrelevant, but it does mean the control belongs at onboarding, recovery and high-risk step-up, not in the everyday login path.

Recovery is where the two meet, and it is the weak point in most consumer stacks. If a sophisticated login is backed by a recovery flow that falls through to a call centre and a few knowledge-based questions, that is where the attack goes. This is one of the day-2 problems that only shows up after a rollout is live.

6. Where Passkeys stop helping#

The most useful session for a passkey-first team was useful precisely because it opened by conceding the point. First slide: passkeys are transforming identity verification. Second slide: sometimes passkeys can't help.

The five gaps named are the ones a passkey rollout does not close:

• Recruitment interviews. No credential is being bypassed, because none has been issued yet.
• Account credential reset. The moment the cryptographic binding is deliberately broken.
• Access to shared devices. Where device binding stops meaning person binding.
• Deliberate credential sharing. A passkey handed over willingly is not an attack the protocol can see.
• AI agents. Software acting for a person, with no face and no device of its own.

The three figures anchoring those:

The last one is the argument for passkeys stated by someone who sells something else. The first two are the argument for why passkeys alone are not a fraud strategy.

6.1 Genuine human Presence as the Root of Trust#

The proposal: when a credential cannot be trusted to represent a person, the fallback has to be evidence of a live human being, independent of any device. Liveness detection is expected to move from incremental to essential to foundational, underpinned by a stack of standards arriving at once: NIST SP 800-63-4, CEN/TS 18099, ISO 25456, FIDO work and eIDAS 2.

Whether liveness can carry that load is open, and it was raised as such on stage. The structure of the argument holds regardless of vendor: a passkey proves the same device and the same key are back. It does not prove the person behind it is who they were at enrollment, and recovery is exactly where that distinction becomes expensive.

7. Newest Pattern: Fraud that gets hired#

The trend flagged as newest is not consumer fraud at all. It is external actors using the same synthetic-identity techniques to pass a remote hiring process:

This is well documented outside the conference. CrowdStrike's 2025 Threat Hunting Report counted more than 320 North Korean IT worker cases in twelve months, close to one investigation per day, and a 220% year-on-year increase in the activity it tracks as Famous Chollima. Generative AI was used at every stage: drafting resumes, building false identities and masking appearance during video interviews. The UN has estimated the scheme generates between $250 million and$600 million per year.

It is an identity problem rather than an HR problem because the attacker ends with legitimate credentials, issued by a legitimate process. No authentication control catches that, since nothing is being bypassed. The controls that apply are verification at onboarding, least privilege afterwards and behavioural detection.

It now reaches fraud vendors as the top enterprise concern from CISOs, and it extends past employees to contractors and third-party vendors. It is also the same attribution problem the agent track spent three days on, with a human in the seat.

8. Fight AI with AI, but keep the Fundamentals#

Defenders need AI to keep pace with the volume and quality of AI-generated attacks, and that retires nothing that already works. The signals the fraud track kept returning to are unglamorous: device, behaviour, name, address, date of birth. The AI layer processes those at machine speed and separates good patterns from risky ones.

The point that gets too little attention is that no single company sees enough of the picture. Fraudsters are not industry-specific, and a profile assembled from a healthcare breach gets monetised somewhere else entirely.

9. What to take back to a CIAM Team#

Four things carry directly into a consumer authentication program.

Stop applying one workflow to everyone. The default path should be the low-friction, phishing-resistant one. Step-up should be a response to a signal, not a policy applied to the whole population. A single mandatory OTP for every user is expensive in both SMS costs and abandoned sessions, and the burden falls overwhelmingly on genuine users who never needed it.

Instrument the whole journey, not just the login endpoint. Signup, login, recovery and high-value transactions are the same identity, and treating them as separate systems creates the blind spot. Seeing that path end to end, across every method a user actually touches, is the case for authentication observability.

Put deepfake controls where deepfakes attack. Verification, recovery and high-risk step-up. Hardening those is worth more than adding liveness checks to a routine login.

Fix the fallback before it becomes the attack path. Every recovery route that reverts to a phone call, an OTP or a security question is where a well-defended login gets bypassed. This is the specific failure mode we cover in passkey day-2 problems.

10. How Corbado can help#

Layered authentication only works if the bottom rung is both low-friction and phishing-resistant. That is the rung Corbado Connect builds: passkeys as the default path in an existing IdP or CIAM stack, so step-up can be reserved for the sessions that actually warrant it instead of applied to everyone.

• Passkey intelligence. Decide per device and per user whether to offer, nudge or suppress a passkey, so the low-friction path is only shown where it will actually succeed.
• Gradual rollout. Rollout rules and A/B tests let you move traffic onto the passkey path in controlled steps rather than switching a population over at once.
• Recovery and fallback in view. Passkey management and fallback flows are part of the same instrumented journey, which is where the weakest handoff usually sits.
• Every step recorded. Corbado Observe captures login, signup and recovery as structured events, so the friction you removed and the friction you kept are both visible per device, browser and market.

11. Conclusion#

The agentic sessions got the attention. The fraud track had the immediately actionable advice, and its message was deliberately unexciting: AI has made existing fraud cheaper and faster rather than replacing it, so the response is to assess risk continuously, apply friction proportionally and stop pretending a single authentication method fits every user and every transaction.

The organisations described as mature were not the ones with the most tools. They were the ones that had broken down internal silos and started looking at identity across the whole customer lifecycle. Any team hits the same conclusion after their first year of running passkeys at scale.

About Corbado

Corbado is the Authentication Intelligence Platform for CIAM teams running consumer authentication at scale. We help you see what IDP logs and generic analytics tools can't: where passkeys, passwords, OTP, social login and fallback journeys succeed, stall or fail, which devices and browsers create friction, and when an OS update silently breaks login. Two products: Corbado Observe layers process mining and observability across authentication journeys. Corbado Connect adds managed passkeys with analytics built in alongside your IDP. VicRoads runs passkeys for 5M+ users with Corbado (+80% passkey activation). Talk to a Passkey Expert →

Frequently Asked Questions#

What is layered authentication?#

Layered authentication assesses risk passively before asking the user for anything, then applies additional verification only in proportion to the risk of that session and transaction. At Identiverse 2026 the sequence was described as device and behavioural assessment first, traditional identity signals such as name, address and date of birth alongside them and step-up authentication only where an anomaly or a high-risk transaction warrants it.

How is AI changing identity fraud?#

Mostly by making existing fraud faster and cheaper rather than by inventing new categories. AI was described at Identiverse 2026 as a force multiplier on schemes that have existed for five to ten years, alongside a smaller set of genuinely new attacks such as deepfakes. The practical implication is that the response is largely better execution of known controls at greater scale, not a replacement architecture.

Do deepfakes break passkeys?#

No. Deepfakes attack identity verification, where an organisation is trying to establish who an unknown person is, typically at onboarding or account recovery. Passkeys are scoped to the domain they were created for and authenticate a returning user cryptographically, so a synthetic face or voice on camera is irrelevant to the login itself. That holds for synced passkeys, which are available across a user's devices, as well as for device-bound ones. The exposure is in recovery flows that fall back to a call centre, an OTP or knowledge-based questions.

How common are deepfake attacks against organisations?#

A Gartner survey of 302 cybersecurity leaders, published September 2025, found 62% had experienced a deepfake attack in the prior 12 months. A separate Gartner CISO survey of 297 respondents found 41% had experienced a deepfake combined with social engineering on an audio call and 35% on a video call. Entrust's 2026 Identity Fraud Report links deepfakes to 1 in 5 biometric fraud attempts globally.

Why do fraudsters target insurance and healthcare?#

For data rather than direct financial gain. Sessions at Identiverse 2026 explained that insurers and health providers hold rich, non-public consumer data, and that fraudsters use access to those systems to complete stolen identity profiles which are then monetised in other industries such as financial services.

What is the fake employee fraud trend?#

It is the use of synthetic identities and AI-altered appearance to pass a remote hiring process and obtain legitimate internal system access. CrowdStrike's 2025 Threat Hunting Report counted more than 320 North Korean IT worker cases in twelve months and a 220% year-on-year increase in that activity, with generative AI used to draft resumes and modify appearance during interviews. It is difficult to catch with authentication controls because nothing is being bypassed, so the defences are verification at onboarding, least privilege and behavioural detection.

Should every user get the same authentication flow?#

No. Applying one workflow uniformly taxes the large majority of genuine users to catch a small percentage of fraud. The layered approach described at Identiverse 2026 assesses risk up front and then selects the workflow, so that recognised devices and low-risk transactions proceed without friction while anomalies and high-value actions trigger stronger verification.