Tamas Kadar

Across case posts, the pattern is the same: accounts look unrelated on the obvious identifiers, then several smaller anomalies line up and reveal one operation behind them. The device or bot's own characteristics, not its stated identity, carry the proof.

A ring of 200 accounts claimed the same sign-up reward while sharing no device ID or IP, so ordinary linking rules saw 200 strangers.

• SEON fraud consultant Kyle Truax, whose work Kadar credits, found no human hesitation during onboarding, identical font counts across every device, one operating-system build ID appearing across different phone models, and only two audio fingerprints for all 200 phones.
• On a marketplace, 5,000 fake accounts running on real Android devices all reported the same battery level, used an external keyboard and shared one Wi-Fi network, leading to a cloud device farm rented out with robotic process automation.
• A bot farm behind 21,000 accounts was first flagged because none of the machines had a microphone or camera connected.

Kadar describes researching fraud in open dark-web forums in 2016 and notes those forums have largely moved to invite-only groups. The detection method has shifted with them, from reading posted tactics to reading device and session signals directly.

Read original · Original 2 · Original 3

Blocking agents also blocks the customers behind them; admitting agents opens the door to fraudsters travelling with them. The third path Kadar advocates ties an agentic session to a known customer, issues a token that cannot be reused or stolen, and then keeps scoring trust at every step, adding friction only when good and bad cannot be told apart.

Amazon blocked Meta's Muse, citing agent verification and stored login credentials, while Walmart, Sephora and Shopify opened their checkouts to it within days. Kadar calls the third approach continuous trust and says SEON is building it.

A year before the later conference recap, Kadar says SEON's customers answered that they would let agents transact as long as the agents made money.

• By Money20/20 the picture was more divided, with Kadar noting no industry consensus on whether to admit agents, let alone who is liable when one buys something the customer did not want.
• Jefferies found a third of tested sites blocking agents.

Read original · Original 2

The mechanism they describe is dynamic friction: make things easier for high-value customers and harder for suspected fraudsters, rather than applying the same steps to everyone. The same continuous-scoring idea carries into the agentic commerce argument, where friction is added only when the system cannot distinguish good from bad.

Kadar's worked example is a customer who signs up on the newest iPhone, in a wealthy ZIP code, with an email address already registered on luxury retail sites.

• That combination suggests a valuable customer, so they can be routed to fewer steps, better offers and VIP treatment from the first session.
• SEON's 1,100-plus fraud signals are presented as the shared basis for both fraud scores and predicted customer value.

Read original · Original 2

Accounts opened by real people with their own documents pass KYC and are then handed over; deepfakes now clear biometric checks that once caught them; authorized push payment fraud involves no break-in at all because the victim sends the money. A related tension Kadar reports is that products ship faster than risk teams can write rules for them.

Mule accounts are opened by real people, sometimes for as little as $30, and sold on after approval so KYC has already signed off.

• Kadar also cites faked selfies rising 58%, AI deepfakes making up one in five biometric fraud attempts, and authorized push payment fraud costing UK banks £576M in a year.
• At Money20/20, a head of risk at a buy-now-pay-later provider said their new wallet was the best product decision of the year, and that risk learned about it when the first losses hit.

Read original · Original 2

The company posts carry the same structured, list-driven style as the fraud posts, and the hackathon is presented as product research rather than only team culture.

Kadar's seven hiring rules, drawn from hiring more than 400 people, include hiring locally for the first ten staff, using cold LinkedIn messages, avoiding the big-logo trap and watching for defensiveness. An annual engineering and product hackathon set two themes, automating internal workflows with AI and improving a customer-facing product with AI, and the winning team built a tool that turns a plain-English description of a fraud workflow into editable dynamic-friction workflows, cutting setup from about an hour to a minute.

Read original · Original 2 · Original 3

Newcomers should start here because it turns abstract fraud detection into a concrete, signal-by-signal investigation.

Tamas Kadar describes how SEON Fraud Consultant Kyle Truax identified 200 accounts claiming the same sign-up reward.

• The accounts did not share a device ID or IP, so ordinary linking rules saw 200 strangers.
• Four signals exposed the ring: no human hesitation during onboarding; the same font count across every device; one OS build ID across different phone models; and only two audio fingerprints for 200 phones.
• Kadar explains each signal in plain terms: real people mistype and pause, font counts vary with apps and language packs, build IDs belong to specific models, and audio fingerprints usually differ across hardware and software combinations.
• Individually each could be coincidence; together they pointed to one script or setup running many times.
• The post ties the case to SEON's Session Monitoring and Network Detection.

For brands planning a campaign

For a webinar or co-authored piece on agentic commerce, this is a ready-made case for why verification must combine behavioral, device, and network signals rather than rely on one shared identifier. It gives a practical example of continuous trust logic: catch anomalies across a session, not just at sign-up.

Start here for a clear example of why single signals fail and context matters.

Tamas Kadar describes catching 5,000 fake accounts on a large streaming-based marketplace.

• The accounts ran on real Android devices with clean histories, so they avoided emulator checks.
• SEON's device data showed every account had the same battery level, an external keyboard connection, one unusual Android setting, and the same Wi-Fi network.
• The trail led to a Chinese cloud device farm that racks Android logic boards and rents them out as a service; an AI layer runs robotic process automation so renters can prompt in plain language.
• Kadar says fraudsters can change individual signals like battery level, so SEON reverse-engineered the farm's setup and shipped an SDK flag, took down five similar farms, and monitors for new ones.
• The post links this to SEON's expansion past 1,100 high-risk fraud signals.

For brands planning a campaign

This is a strong proof point for a session on AI shopping agents: it shows that agents or bots can pass isolated checks while still sharing infrastructure-level tells. A campaign angle could contrast single-point verification with layered device and network analysis, especially for merchants deciding whether to block or verify agent traffic.

Newcomers should start here for origin story and a concrete bot-farm example.

Tamas Kadar recounts running a crypto exchange with Bence Jendruszak from a university apartment in 2016.

• They processed $10,000 in their first week, then scammers found them and chargebacks began.
• Fraud vendors turned them away because they only worked with big companies.
• Kadar and Jendruszak researched dark web forums where fraudsters shared tactics, staying up all night reading tutorials.
• Kadar notes those forums have mostly moved to invite-only groups, some charging up to $1,000, and effective methods are rarely shared widely.
• The post also describes catching a bot farm behind 21,000 accounts by first noticing none of the computers had a microphone or camera connected, then confirming with other signals.
• Kadar contrasts that with SEON checking 1,100+ fraud signals per new account.

For brands planning a campaign

For a brand or agency considering Tamas Kadar, this post shows narrative range: personal origin story, fraud-economy context, and a memorable detection example. It can open a webinar or co-authored piece by grounding verification strategy in why single checks fail and why layered signals became the norm.

Start here because it frames the live commercial decision around agentic commerce.

Tamas Kadar says Meta's Muse and OpenAI's Dots can now shop for users, Amazon blocks Muse, and Shopify lets it in.

• Kadar outlines three options for companies.
• First, block agents outright to prevent fraud, as Amazon did, citing agent verification and stored login credentials, though this also blocks the customers behind them.
• Second, let agents in to drive revenue, as Walmart, Sephora, and Shopify did within days of Amazon's block, while risking fraudsters hiding among good agent traffic.
• Third, verify once, tie an agentic session to a customer, give the agent a token nobody else can use or steal, then keep scoring trust at every step and add friction only when good and bad cannot be separated.
• Kadar calls this continuous trust and says SEON is building it.
• The post links to a Fintech Business Weekly conversation with Jason Mikula.

For brands planning a campaign

This is the most direct fit for the reader brief.

• It supplies a ready-made webinar structure: frame the block, allow, or verify choice, then go deeper on continuous trust for AI shopping agents.
• A co-authored piece could compare the three paths and explain what verification data a merchant needs before letting agents transact.

Start here for a concise industry snapshot that connects agentic commerce to current fraud operations.

Tamas Kadar reports three reflections from Money20/20.

• First, the industry cannot agree on AI agents or liability: Walmart and Sephora embraced them, Amazon blocked Meta's Muse, and Jefferies found a third of tested sites did the same.
• There is no consensus on who is responsible if an agent buys something the customer did not want.
• Second, products ship faster than fraud teams can keep up; a BNPL head of risk said their new wallet was the best product decision of the year, while risk learned about it when losses hit.
• Third, fraud has moved to where identity checks do not look: fraudsters pay real people as little as $30 to open accounts with their own ID and face, then hand over the approved mule account, so KYC has already signed off.
• The post thanks attendees and links to a recap.

For brands planning a campaign

This post helps position Tamas Kadar as someone who can speak to both agent strategy and operational reality. A campaign could use it to shape a webinar around unanswered liability questions, then bring in SEON's verification approach as a way to reduce uncertainty before agent traffic scales.

Start here to see the commercial argument a brand would care about.

Tamas Kadar argues fraud teams are often seen as revenue blockers but sit on data that can predict spending.

• If a company can predict who will steal, it can predict who will spend.
• Factors like email age, phone registration, and sign-up device help predict customer value.
• Kadar gives an example: someone signing up on the newest iPhone, in a wealthy ZIP code, with an email registered on luxury retail sites is likely valuable, so they can get fewer steps, better offers, and VIP treatment from the first session.
• SEON uses the same 1,100+ signals behind fraud scores to estimate customer lifetime value.
• The post frames dynamic friction as a way to catch fraud and drive revenue at the same time.

For brands planning a campaign

For a brand or agency, this broadens the pitch beyond fraud prevention: Tamas Kadar can connect risk signals to revenue and customer experience. It is a useful secondary campaign angle for a webinar or co-authored piece, especially if the brief needs a commercial rationale for investing in verification before AI agent traffic grows.

On a rewards platform, 200 accounts claiming the same sign-up incentive looked unrelated because they shared no device ID or IP, but all showed no human hesitation during onboarding, identical font counts, one OS build ID across different phone models, and only two audio fingerprints.

• On a streaming-based marketplace, 5,000 fake accounts all reported the same battery level, ran through an external keyboard, and traced back to a Chinese cloud device farm running robotic process automation.
• A separate bot farm behind 21,000 accounts was first flagged because none of the machines had a microphone or camera connected.
• The recurring argument across all three is that any individual check can be passed, so detection depends on many signals read together in context.

Each ring is told as a numbered list of the tell-tale signals, each signal explained in plain language and then connected to the SEON product built for it, such as Session Monitoring or Network Detection.

Kadar credits colleagues for the underlying discoveries in at least one case, and the posts name the fraud consultant who found the signals, so attribution should follow the original credit rather than being reassigned to Kadar alone.

For brands planning a campaign

These cases give a brand ready-made, inspectable examples for a webinar or article on layering identity, device, and behavioural signals. An audience can test the same four or five signals against its own onboarding flow.

Examples: Source 1 · Source 2 · Source 3

A later conference recap describes an industry that cannot agree on whether to admit agents, let alone who is liable when an agent buys the wrong thing, with a third of tested sites blocking agents. The argument is that blanket blocking shuts out real customers behind the agents.

The argument is built option by option, each option paired with named companies that chose it, then a proposed third path with a product mechanism attached.

This is a fast-moving area, and the download figures and merchant decisions cited come from the posts themselves. A campaign would need to refresh those specifics before publication.

For brands planning a campaign

This is the closest match to the reader brief on agentic commerce verification. Kadar can carry a webinar or co-authored piece on verifying and scoring AI shopping agents, using the block, admit, or verify framing and concrete merchant examples.

Examples: Source 1 · Source 2

For an October 15 panel with Visa, Prove, and FinTech Futures, the three covered were deepfakes that beat KYC checks, where faked selfies rose 58% and AI deepfakes make up one in five biometric fraud attempts; mule accounts bought after they have passed KYC, opened by real people on platforms including Telegram and then sold on; and authorized push payment fraud, where the victim sends the money and which cost UK banks GBP 576M in a year. The Money20/20 recap returns to mule accounts, describing fraudsters paying people as little as USD 30 to open accounts with their own ID and face, after which KYC has already signed off.

Each fraud type gets a plain-language definition, a mechanism, and a scale figure, structured as a three-part talk outline rather than a loose list.

For brands planning a campaign

The material is already formatted as speaker content, so it suits a sponsored panel, a co-hosted webinar, or expert commentary on emerging fraud typologies. The mule-account angle is repeated across two events, which suggests it is a durable talking point.

Examples: Source 1 · Source 2

Signals such as email age, phone registration, and sign-up device can route high-value customers to fewer steps and better offers while adding friction only where the system cannot tell good from bad.

• The same continuous-scoring idea appears in the agentic commerce discussion, where trust is scored per session and friction is added only when needed.
• SEON's 1,100-plus fraud signals are presented as the shared basis for both fraud scores and predicted customer value.

The argument is built from a worked example, such as a customer signing up on the newest iPhone in a wealthy zip code with an email registered on luxury retail sites, then generalised into a routing rule.

For brands planning a campaign

This gives a commercial hook for campaigns aimed at risk and fraud leaders who are asked to justify their budgets. It supports a co-authored piece on using fraud infrastructure for revenue decisions rather than only loss prevention.

Examples: Source 1 · Source 2

From Money20/20: no consensus on AI agent liability, products shipping faster than risk teams can write rules for them, and fraud moving to where identity checks do not look. From RiskTech Connect with 86 fraud and AML leaders in New York and Berlin: most companies want AI to support rather than replace analysts, a new fraud pattern takes 30 to 60 days to detect while most attendees guessed under 14 days, half of fraud teams run AI models they do not fully trust, and proving business impact is now the top challenge.

Both posts use a short numbered list of findings, presented as what leaders in the room said rather than as vendor claims, and each closes with an invitation to the next event.

For brands planning a campaign

The format suits a co-hosted roundtable, research report, or webinar where a brand wants audience-sourced findings rather than a straight product pitch. The 30-to-60-day detection figure is a specific, quotable claim a campaign could build an agenda around.

Examples: Source 1 · Source 2

One post gives seven hiring rules drawn from hiring 400-plus people, including hiring locally for the first ten staff, using cold LinkedIn messages, watching for defensiveness, and avoiding the big-logo trap, with named influences from Revolut's hiring playbooks and the book Who by Geoff Smart and Randy Street. Another describes an annual engineering and product hackathon held in a European gaming centre, with teams working on automating internal workflows and improving customer-facing products with AI, and a winning tool that turns a plain-English description of a fraud workflow into editable dynamic-friction workflows.

The hiring post uses a numbered rules list; the hackathon post names the themes, the winning team members, and the demo outcome.

For brands planning a campaign

This is a leadership and employer-brand angle rather than a core fraud brief. It could support a broader founder-profile piece or a culture segment, but a fraud-focused campaign will find more direct material in the case-study posts.

Examples: Source 1 · Source 2

For Y Combinator, SEON screens applicants on the Work at a Startup hiring platform used by more than 5,000 portfolio companies, where fraudsters submit synthetic AI profiles and fabricated emails to collect pay under a false identity or phish founders. The post lists three ways SEON helps: tracking the applicant journey, providing an intelligence layer for YC's own models, and feeding every decision back as training data through a labelling API.

The partnership is told as a problem, a vendor-selection note, and three numbered capabilities, ending with a link to more detail.

For brands planning a campaign

The post shows Kadar will co-produce content built around customer evidence and a named partner. It is a useful model for a co-authored campaign piece where a brand supplies the platform and SEON supplies the detection layer.

Examples: Source 1 · Source 2

Turned away by established fraud vendors, they spent the summer reading dark-web forums where fraudsters shared methods. The post then describes how those forums moved to invite-only groups charging up to USD 1,000 to join, and how detection now works from device signals such as a bot farm behind 21,000 accounts whose machines had no microphone or camera.

A narrative origin story that moves from the founding problem to the research method to the current product, using one past and one present example.

For brands planning a campaign

The story works as a human opener for a podcast, long-form profile, or keynote introduction, and it gives Kadar a firsthand research narrative that supports credibility claims in a campaign.

Examples: Source 1

The posts refer to SEON's fraud signals, scoring, and risk products, and one post says Kadar has hired 400+ people at SEON. The material does not state founder or owner status.

View example

The post says YC chose SEON after evaluating multiple vendors, and describes SEON tracking applicant signals, supplying data for YC's custom models, and feeding decisions back through a labeling API. The relationship is presented as a SEON customer story.

View example

Machado explains how selling through Meta's Muse lands in Shopify and how ChatGPT Ads connect to HubSpot and Shopify, which is the commercial side of the agentic commerce question Kadar approaches from risk.

Machado stays with marketing and channel execution, such as post-purchase flows and ad tooling, while Kadar works on verification, fraud scoring, and liability when an agent buys the wrong thing.

Read original · Original 2

Both run a comparison of named agents against what merchants have made readable to them.

Willoughby frames the problem around whether a brand's own knowledge is legible to agents and what that does to discoverability, while Kadar frames it around verifying the agent and scoring trust through the session.

Read original · Original 2

Moock is an infrastructure and platform-distribution view, focused on making catalogs agent-ready through one integration, while Kadar is a detection view, focused on telling good agents and customers from fraudulent ones.

Read original · Original 2

The posts worked through the operational consequences of letting Meta's Muse sell for a Shopify store: the order still lands in Shopify, so a connected email platform can still trigger post-purchase flows, while website analytics pixels do not fire during the agent checkout.

• A companion post covered connecting ChatGPT ads to HubSpot and Shopify and OpenAI testing conversational ads.
• That is the enabling-side view: how a merchant makes agent commerce work commercially and keeps follow-up working.

The material is about opening the door, not about who walks through it.

• There is no device or identity signal analysis, no fraud-ring case study and no trust-scoring or token mechanism.
• A fraud-prevention audience would get the merchant experience of agentic commerce but not the verification depth the brief is built around.
• Choosing this alternative buys marketing-stack fluency at the cost of the detection detail the campaign was planned around.

Examples: Source 1 · Source 2

The posts supply a ready-made exercise — ask one of the current personal agents to buy your own product and watch what it reads — and the argument that an agent works only from what a brand has put where software can read it, filling gaps with the average when product pages, reviews and claims are thin.

• A second post brought a benchmark of LLM agents running a simulated store for a year, where the best earned 27.3% of what human beginners did, and adding a memory-and-skills harness lifted results.
• That supports a session about encoding a company's specific knowledge where an agent can act on it.

The lens is agent readiness and system design, not fraud.

• There is no identity verification, no signal inventory, no scoring or token mechanism, and no case study of stopping a fraudulent account.
• A verification-strategy session would need a fraud specialist alongside this speaker.
• Choosing this alternative gets a strong merchant-facing demo and a strategic frame about what agents can and cannot know, and gives up the detection and trust-scoring core of the brief.

Examples: Source 1 · Source 2

The posts explained a single-integration model that makes an entire catalogue discoverable to a range of AI agents, with checkout, payments and fraud protection bundled, and listed platform and brand early adopters by name. A second post read OpenAI's pullback of checkout inside ChatGPT as evidence that discovery matters more than a native buy button and that open, agent-agnostic infrastructure wins long term.

The account is close to the product it describes and the posts read as launch material for that suite, so a session would carry that vendor's framing.

• Fraud protection appears as a built-in feature rather than an explained strategy, so the specific verification questions in the brief — how to verify an agent once, how to tie a session to a customer, how to keep scoring trust — would go largely unanswered.
• Choosing this alternative buys platform-architecture fluency and a multi-agent road map at the cost of the independent detection evidence a fraud-prevention brand is looking for.

Examples: Source 1 · Source 2