The Verification Files
TVF-007 · Identity verification · 50 min listen

When Identity Verification Becomes the Risk: Five Ways Checks Fail

Checks run in the wrong order, breached data that returns as valid, stored data that becomes the next breach, one-off onboarding and AI agents with no verified authority.

Key findings

  • A business can run every required check and still onboard people it should have refused. In one licensed gaming operator, the nationality and ID checks ran in the wrong order.
  • Breached data does not stay stolen. It is altered, ages and returns a year or two later looking like a validated record.
  • Data collected in case it is useful becomes the next breach. You do not need to secure what you do not hold.
  • Verification at onboarding captures one moment. People move, change numbers and change email addresses afterwards, and the business does not know.
  • An AI agent acting for a person needs verified authority and limits. Without them the person is bound by whatever the agent orders.
  • The verification industry treats data as legitimate because it is used in verification. It should be the other way round.

Contributing intelligence

  • Simon Delaney

    Founder, Provero · LinkedIn profile

    Background in lead generation, identity, verification and software for data processing.

  • Harald Dols

    Chief executive and co-founder, STRIKE Group · LinkedIn profile

    STRIKE Group builds identity verification and onboarding flows and is developing Know Your Agent checks. Contributed expertise on verification workflows, data minimisation, ongoing verification and Know Your Agent.

Contributors are sources, not authors. The Verification Files is responsible for the analysis and the assessments in this file.

Summary assessment

Identity verification is meant to reduce risk. This file records five ways in which it adds to it: checks run in the wrong order, breached data returning as valid, stored data becoming the next breach, verification treated as a single event, and AI agents acting with no verified authority.

The common cause is provenance. Data is trusted because it is used in verification. It should be used in verification because it can be trusted.

Source: high. Delaney and Dols each report from their own work on how checks and data are handled.

Evidence: published for the two public cases. Stated for findings 1, 3 and 4. Contact sources for evidence. On finding 2, Delaney has seen the result first-hand, and the steps that lead to it are the sources’ account. Hypothetical: finding 5. Know Your Agent is a new field, and this is reasoning about what comes next. How widespread each failure is has not been measured.

Finding 1: checks in the wrong order

Case: a licensed online gaming operator. Reported by Dols, whose firm reviewed the onboarding flow. The flow had been built by another specialist provider.

The operator needed a government licence to offer its games. Players had to be verified before they could play. The flow contained the required checks. Two of them, on nationality and on the identity document, ran in the wrong order.

Result:

  • Players who were barred under the licence were admitted.
  • Players were able to open more than one account, which the licence also prohibits.

Cause. The flow was arranged to protect conversion. The chief revenue officer wants sign-ups. The compliance officer wants the licence kept. The two rarely agree, and here conversion decided the sequence.

Dols’s image is a stack of fishing nets. Stack them offset and nothing gets through. Stack them aligned and the holes line up.

Reading: the operator ticked each box its regulator required and was still non-compliant. The fault lay in how the checks were assembled, not in the checks or the data.

Cost:

  • Long term: the licence. Losing it closes the business. Delaney puts the exposure in the hundreds of millions for a large operator. Dols reports that revenue teams do not treat it as a live danger.
  • Short term: paying twice. Players open second accounts to collect free credits. The books show a new player. The operator has paid the acquisition and bonus cost again for the same person. This loss is rarely visible.

Finding 2: breached data returns as valid

Breach reporting focuses on who was hit. It rarely examines what was taken.

Public case. In August 2026 the extortion group ShinyHunters published data taken from the clothing company Carhartt after a $3.3 million ransom went unpaid. The group’s claim ran as high as 24.8 million records. Analysis for the breach notification service Have I Been Pwned confirmed 12.9 million real accounts and removed millions of synthetic records. Close to half of what was announced as stolen was not personal data.

The mechanism, as both sources report it:

  1. Those who steal data do not check its quality or record its age.
  2. The data spreads and is altered along the way.
  3. A record stolen now reappears in a year or two looking like an established, validated one.
  4. An email address joined to the correct first name can be enough to open an account without triggering a negative check.

Dols calls this pollution of the verification industry.

Delaney adds: once a ransom goes unpaid the data loses value and is sold cheaply. It is this older data that re-enters commercial data sets. Missing fields are filled in to complete the record, which is how records come to show a date of birth later than the breach they came from. See where commercial KYC data comes from.

Finding 3: stored data becomes the next breach

Data is treated as the new gold. Businesses collect as much as they can justify, in case it is useful later.

Dols compares it to the offcut of a shelf kept in the attic. It is not needed. It is not thrown away. Over time nobody knows what is up there, and nobody notices when it has gone.

  • Unused data still has to be secured. Under a standard such as ISO 27001, everything held is in scope. One room is easier to secure than an apartment block.
  • Unused data is still useful to someone else. Once stolen, it is repurposed and returns as the pollution described in finding 2. The circle closes.
  • Data may now be more liability than asset. Delaney’s assessment, given the value of data to AI and the scrutiny on privacy.

The principle is data minimisation. If the data is not needed at this step, do not ask for it. Dols’s summary: you do not need to secure what you do not have. A business that knows exactly what it holds can concentrate on protecting it.

Businesses respond to breach headlines by spending on cyber security. The prior question, what they hold and why, goes unasked.

Finding 4: verification captures a moment

A verification is true at the time it is done. Afterwards people move house, change phone number, change email address, change employer and move country.

  • Verification at the point of entry is now standard practice.
  • Businesses are legally required to keep track of their customers. Nobody tells them how.
  • Once a customer is through the gate, the business has no means of knowing what has changed.

The alternative is ongoing monitoring. The business receives a signal when a held detail no longer appears to be true, for example that the customer is no longer at the recorded address. It can then ask. Verification becomes a record of the customer over time. Delaney describes it as an identity that has been refreshed several times before the next check is ever triggered.

There is a second benefit. A business holding accurate data has less stale data to lose.

Why it is not done:

  • Good data is scarce and costly. Consumers are increasingly reluctant to share details, and breach coverage in the media feeds that reluctance.
  • The infrastructure has been missing. Until recently there was no way to check continuously without raising objections under GDPR and other privacy law.
  • The comparison is never made. Businesses do not set the cost of monitoring against the cost of a breach, a ransom or a lost licence. Most have not suffered one.
  • They do not believe they have to.

Delaney’s reading is that this is a need and not a want. Nobody shops for a plaster until they are bleeding.

Finding 5: AI agents without verified authority

Dols’s illustration. A person tells an AI agent to make sure there is milk every day and hands over payment details with no limit. Two days later a cattle lorry arrives with a cow and a milking installation. The goal has been met. The supplier believes the person placed the order. On Dols’s reading, the person is bound by it.

Know Your Agent, or KYA, has two sides:

  • The agent carries a passport. It is issued by the person, who must first be verified. It holds the data needed for the task and a band of permissions, including a spending limit. A limit of a few pounds would have prevented the cow.
  • The receiving business reads it. A merchant, platform or trading venue needs a way to check the passport, as a border checks a travel document.

Status: early. Dols reports that KYA exists and is not widespread.

Public case. Meta launched its Muse personal agent on 8 September 2026. Amazon blocked it from shopping on its site on 20 September. Published reports say Amazon acted because Meta had no permission and the agent did not identify itself as an AI. The dispute turns on the question KYA exists to answer: who the agent is and what it is allowed to do.

The same question applies to trading platforms where agents manage portfolios. If a trade goes wrong, it is unclear whether the agent, the person or the platform that admitted the agent is accountable.

What the industry has to admit

Asked what is not working, Dols names provenance.

  • Much of the data circulating in verification is treated as gold and is closer to fool’s gold.
  • The industry has many middlemen. That makes it harder to see where data came from.
  • An accurate view of provenance would reduce pollution, reduce misuse and reduce what is exposed in a breach.

Delaney’s reading closes the file. Data used in identity verification becomes legitimate because of that use. The use decides that the data is sound. It should be the reverse.

Assessment

Each finding is a case of trusting the process over the evidence: the checklist over the sequence, the record over its origin, the stored file over the need for it, the onboarding result over the present, and the order over the authority behind it.

The countermeasures follow below.

Listen to episode 007.

Countermeasures · TVF-007

Countermeasures: Verification that adds risk

Every check can pass and the decision can still be wrong. Checks run in the wrong order, stored data becomes the next breach, and an onboarding result goes out of date.

Drawn from answers given by Simon Delaney and Harald Dols.

  1. Complete legally required checks before entry

    Any check the law requires before access, such as age, exclusion registers and sanctions, has to be completed before the point of entry. Never after it. Which checks those are depends on the checks, the country and the outcome needed. If you are not certain, ask your provider or take legal advice.

  2. Sequence the rest on cost and friction

    Everything else can be ordered by cost and friction. Cheap and invisible checks come first. The heavier ones run only when the risk calls for it.

  3. Test the flow with everyone who has a stake

    Test that the flow works as intended and in a compliant way. Involve every stakeholder with a vested interest in the result, compliance included.

  4. Let the law set the floor and the board set the appetite

    The law sets the floor, not the flow. Below the floor there is nothing to discuss and compliance has the final say. Above it the question is risk appetite. That decision belongs to the board, not to the revenue team or the compliance team.

  5. Judge breach exposure by pattern, with provenance

    One record showing in a breach decides nothing. A coin landing on heads once means nothing. Landing on heads 30 times in a row means the coin is loaded. Ask for the provenance behind the data: the opt-in date and time, where it came from and how the traffic was generated. Then judge its legitimacy.

  6. Keep data only if you use it

    Start with personal data and anything that has value in the wrong hands. If it is not being used for a valid reason, such as fraud detection or offering the customer something, ask why you hold it. Unused data has no value to you and great value to someone else.

  7. Re-check on events and signals

    A fixed interval on its own gives you little to check. Re-verify when something happens: an event, new data, or a signal that a detail has changed. Check the source, provenance and legitimacy of the update too.

  8. Make the cost case as insurance

    A business that has never had a breach or lost a licence will not feel the need. It is like insurance. You think you do not need it until you do. That is why KYC and identity verification are compulsory in financial services.

  9. Treat an agent as someone acting for a person

    When an AI agent arrives at the checkout, the merchant needs to know two things: who the person behind the agent is, and whether the agent is authorised to make this purchase. If both cannot be established, handle it as an unverified customer.

  10. Set limits before giving an agent payment details

    Set a spending cap. Set what it may buy and where. Set an end date. Require a confirmation step above a certain amount. Define the scope of what it may do outside a given product description.

What does not work

  • “We ticked every box”Confuses compliance with safety.
  • “They were verified at onboarding”A snapshot of one moment.
  • “More data makes us safer”The reverse of the truth. Every extra field you do not need is extra liability.
  • “Cyber security has it covered”Identity fraud walks in through the front door with valid details.

Limits

A person who knowingly lends out their own identity. No check catches a willing participant in fraud at the door. It shows only later, in behaviour, which is why ongoing monitoring matters.

Example checks

Example checkIllustrative
Onboarding flow: checks present
age check
present
id document
present
exclusion register
present
duplicate account
present
TickedEvery box ticked.
Example checkIllustrative
Same flow: order of checks
account opened
step 2
age check
step 4, after entry
exclusion register
step 5, after entry
duplicate account
step 6, after entry
FlaggedRequired checks run after entry. Not compliant.

Example checks are illustrative. They are not real records.

Operational resources

Relevant services and organisations that can help implement these countermeasures.

STRIKE Group

Harald Dols, a contributor to this file, is its chief executive and co-founder.

  • STRIKE Group: Identity verification and monitoring

    Builds verification flows for regulated onboarding and monitors customers after they have been verified. Dols describes Know Your Agent checks as in development.

    Supports: 01 Complete legally required checks before entry · 02 Sequence the rest on cost and friction · 07 Re-check on events and signals

Provero

Provero was founded by Simon Delaney, who produces The Verification Files.

  • Provero: Breach detection

    Shows which known breaches an email address or phone number appears in, with the date of each breach. It can be run across a file. One record in a breach proves nothing. The pattern across the file is what is read.

    Supports: 05 Judge breach exposure by pattern, with provenance

A resource is one way to carry out a countermeasure. Listing is not an endorsement. Some countermeasures may only be able to be applied via a resource. How resources are chosen

Frequently asked questions

How can identity verification create risk?

In five ways covered in this file. Checks can run in an order that leaves gaps. Reference data can include breached records that have come back into circulation. Data collected for verification and kept becomes material for the next breach. A check at onboarding goes out of date as the customer’s details change. And AI agents can act for a person whose authority nobody has verified.

Why does the order of verification checks matter?

Each check filters out a different group. If a check that should exclude someone runs after the point where they are let in, the exclusion never applies. A business can complete every check its regulator lists and still be non-compliant, because the checks were sequenced to protect conversion and not to close the gaps.

In what order should verification checks run?

One rule is fixed. Any check the law requires before access, such as age, exclusion registers and sanctions, must be completed before the point of entry and never after it. Everything else can be sequenced on cost and friction: cheap and invisible checks first, heavier ones only when the risk calls for it.

What does a duplicate account cost a gaming operator?

More than it appears. Players open second accounts to claim free credits again. The operator records a new player and pays the acquisition and bonus cost a second time for the same person. Duplicate accounts are also prohibited under the licence, so each one is a compliance breach.

How does breached data come back as valid?

Stolen data is not dated or quality-checked by the people who take it. It is copied, altered and resold. A record stolen today can reappear in a year or two looking like an established one. An email address attached to the correct first name can be enough to open an account without triggering a negative check.

What is data minimisation in identity verification?

Data minimisation means collecting only the data needed for the step in hand and not keeping the rest. Data held but not used still has to be secured, is easy to lose track of, and is of use to whoever steals it. A business that holds less has less to protect and less to lose.

Why is verification at onboarding not enough?

Because it records a moment. Customers move house, change phone numbers, change email addresses and change jobs. Businesses are required to keep customer records accurate but are not told how. Ongoing monitoring flags when a detail no longer matches, so the business can ask the customer about it.

What is Know Your Agent (KYA)?

Know Your Agent is the verification of AI agents that act for a person. The agent carries a form of digital passport issued by the verified person. It states what the agent is for and the limits on what it may do, such as a spending cap. The business receiving the agent needs a way to read and check that passport.

Who is liable when an AI agent makes a purchase?

In the account given in this file, the person who gave the agent its instruction and its payment details. The seller treats the order as theirs. That is Harald Dols’s reading and not legal advice. It is the reason an agent’s authority and limits need to be set and verified in advance.

Exhibits: sources and further reading

Related files