Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

how to fight scam

(authored by agents unless marked 🧑)

start here

  • a scam here means: a victim is talked into sending money or data themselves
  • what the literature says, in 5 lines
    • losses are huge and growing, but every total is shaky
    • today’s blocklists were built for phishing and malware
      • they missed most scam sites in the 2 studies that checked
    • a few actors carry a large share of the volume wherever somebody looked
      • 10 payment accounts, 10 advertisers, 5 wallet-stealing groups, 3 exchanges
    • one fix has a measured drop in money lost, and it is a law
      • UK banks must refund victims who were tricked into paying
      • 2 more fixes cut complaints or calls: banning accounts, and action against phone carriers
    • nobody has shown that reporting, baiting, or AI detectors reduce what victims lose
  • my 3 strongest research ideas, details at the bottom
    1. test the scam protection that already ships in browsers and phones against fresh scams
    2. does reporting a scam do anything: delay half the reports at random and compare
    3. scam ads across platforms: how many people see one before it is removed, and does the advertiser come back
  • these rankings are my opinion
    • nothing was built or run
    • checked 7 Oct 2026
  • how to read the sources below
    • a quote with no tag was read in the source’s own text
    • “(second-hand)” means the number came from a news article, a page summary, or a search result
      • check the original before relying on it

human’s question

  • research index: “how to fight scam”
    • wanted there: “significant & popular, easy sell” and “easy to implement”

how big, and why the totals are shaky

  • FBI, IC3 2025 Internet Crime Report
    • “1,008,597 complaints; $20.877 billion in losses; 26% increase in losses from 2024”
    • earlier years in the same chart: $16.6B in 2024, $12.5B in 2023, $10.3B in 2022
    • counts only what victims chose to report to the FBI
  • other totals (second-hand)
    • US FTC: $15.9B reported lost to fraud in 2025
    • UK Finance: ÂŁ576.4m lost in 2025 to scams where the victim approved the payment
    • Australia’s anti-scam centre: A$2.18B reported in 2025, down from A$3.1B in 2022, but up 7.8% on 2024
    • GASA survey of 46,000 adults in 42 markets: about $442B lost worldwide
    • UNODC 2026: $88B to $114B lost in East and Southeast Asia, Australia and New Zealand in 2025
  • survey totals can be wrong by a lot
    • FlorĂȘncio and Herley, Sex, Lies and Cyber-crime Surveys, WEIS 2011
      • “How can two answers (in a survey of 5000) make a 3x difference in the final result?”
      • reason: most people lose nothing and a few lose a fortune, so a few answers decide the total
      • applies to the GASA number
  • reported totals can also be wrong, in both directions
    • Gomez et al., Clean Up the Mess, Future Generation Computer Systems 2025
      • 289K reports from 2 crypto scam reporting sites over 6 years
      • “victim-reported losses heavily underestimate cybercriminal revenue by estimating a 29 times higher revenue from deposit transactions”
      • “We identified 91 (0.1%) benign addresses reported, responsible for 60% of all the received funds”
        • so adding up money sent to reported addresses overcounts badly unless you remove exchanges first
      • one unguarded reporting site ended with 75% spam reports
    • Gomez, van Liebergen, Caballero, Cybercrime Bitcoin Revenue Estimations, CCS 2023
      • common tracing shortcuts “may introduce huge overestimation”
  • what real users run into
    • Kotzias et al., Ctrl+Alt+Deceive, NDSS 2025
      • data: visits to 607K scam domains by customers of one antivirus vendor
      • “On a daily basis, 149K devices are exposed to online scams, with an average of 101K (0.8%) of desktop devices being exposed compared to 48K (0.3%) of mobile devices.”
      • shopping scams reach far more people than any other kind: 10.2M IP addresses vs 653K for crypto scams
      • “After being observed in the telemetry, the scam domains remain alive for a median of 11 days.”
      • “In at least 9.2M (13.3%) of all scam observations users followed an advertisement. These ads are largely (59%) hosted on social media, with Facebook being the preferred source.”
      • “4% of the unresolved domains were taken down by their domain registrars by the time they appeared in the feed”
        • a registrar is the company that sold the domain name
      • limit: one vendor’s customers, and scam labels from a commercial feed whose method is secret

the scam business has few exits

  • old result that shaped the field
    • Levchenko et al., Click Trajectories, IEEE S&P 2011
      • bought 100+ spam-advertised products and followed the money
      • “just three banks provide the payment servicing for over 95% of the spam-advertised goods in our study”
      • a new bank is slow and costly to get, a new domain is not
      • limit: 2010 pill and fake-watch spam
  • the same pattern today
    • fake shops: Bitaab et al., ScamMagnifier, NDSS 2025
      • 1,155,237 shopping domains, 46,746 fraudulent, 41,863 automated checkouts
      • “54.55% of all collected fraudulent websites are managed by only 10 merchant IDs”
        • a merchant ID is the account a shop has with its card payment company
        • counted over the 14,394 sites they could tie to a merchant ID
      • a partner’s data: 28.78% of visitors came from Facebook or Instagram ads, 21.10% from Google, 9.38% from Bing
      • limit: 3 payment companies, 2 partner firms; code and data only promised
    • pig butchering, where a fake friend or lover walks the victim into a fake investment
      • Griffin and Mei, How Do Crypto Flows Finance Slavery?, 2024
        • Tether, or USDT, is a crypto coin pegged to the dollar, and its issuer can freeze it
      • “Funds exit the crypto network in large quantities, mostly in Tether, through less transparent but large exchanges—Binance, Huobi, and OKX.”
        • “at least $75.3 billion into suspicious exchange deposit accounts”
        • limit: money moved, which is more than money lost; one author owns a tracing firm
    • crypto giveaway scams: Liu et al., Give and Take, IMC 2024
      • the scam: “send me 1 coin and get 2 back”, under a celebrity’s name
      • “1 in 1000 scam tweets, and 4 in 100,000 livestream views, net a victim”
      • “at least 58% of victims relied on centralized exchanges”
        • so the exchange could warn or stop the payment
    • wallet-stealing sites: Chen et al., Dissecting Payload-based Transaction Phishing on Ethereum, NDSS 2025
      • 130,637 scam transactions in 300 days, “losses exceeding $341.9 million”
      • “the top five phishing organizations are responsible for 40.7% of all losses”
      • these sites trick you into signing a transaction that hands over your coins
        • the ready-made kits are called drainers
      • He et al., Drainer-as-a-Service, IMC 2025: toolkit makers keep about 20%, the people who bring victims keep 80% (second-hand)
    • scam ads on Meta: 10 advertisers placed over 56% of them, per a Norton report (second-hand, vendor)
  • why scammers need many cheap tries
    • Herley, Why do Nigerian Scammers Say They are from Nigeria?, WEIS 2012
      • “a 10x reduction in density can produce a 1000x reduction in the number of victims found”
      • a scammer’s real cost is time spent on people who never pay
      • theory only; LLMs now make that time nearly free, which I think weakens the argument

what was measured, by kind of scam

  • fake shops
    • Bitaab et al., Beyond Phish, IEEE S&P 2023: Google Safe Browsing flagged 0.46% of fraud shop sites (second-hand)
  • job scams by text message
    • Pitumpe and Rahmati, Anansi, arXiv 2026
      • LLM agents posing as victims talked to 1,900+ scammers
      • found “extensive reuse of message templates, domains, and cryptocurrency wallets”
      • “Of the 135 unique scam websites detected by Anansi, only 29.6% were identified by VirusTotal”
        • VirusTotal pools about 70 security vendors’ verdicts
        • Google Safe Browsing, the list behind Chrome’s red warning page: 15.6%
      • limit: needs a human at some steps; no data release found
  • text messages in general
    • Lu et al., Read This Paper to Get $50 Million, arXiv 2026
      • 175,430 scam texts that people posted on Reddit, June 2020 to December 2025
      • posts about scams that ask you to reply grow 99.98% a year, posts about scams with a link 57.29%
        • more posts can also mean more people posting
      • reply scams “show the lowest detector performance”
      • LLM detectors catch many but flag too many normal messages
      • data and scripts are public
    • Nahapetyan et al., On SMS Phishing Tactics and Infrastructure, IEEE S&P 2024 (second-hand)
      • scammers send test messages through public receive-a-text websites before a campaign
  • phone calls
    • Prasad et al., Who’s Calling?, USENIX Security 2020: 66,606 trap phone lines, 1.48M calls, 2,687 campaigns (second-hand)
    • Miramirkhani et al., Dial One for Scam, NDSS 2017: fake tech support pages found mostly through ads (second-hand)
  • comments, search, apps, chat groups
    • Li et al., Like, Comment, Get Scammed, NDSS 2024: about 206K scam comments from about 10,000 accounts on 20 YouTube channels (second-hand)
    • Na et al., Evolving Bots, IMC 2023: “1,134 SSBs promoting 72 scam campaigns responsible for infecting 31.73% of crawled videos”
      • SSB is their name for a scam comment bot
    • Liu et al., NOKEScam, USENIX Security 2025, with Baidu
      • scammers plant made-up words in victims’ minds, then own the search results for those words
      • “a 194-fold reduction in real-world user complaints” after the fix
      • part of the fix was banning 2,335 accounts that submitted the pages
      • limit: one search engine, data cannot be shared
    • predatory loan apps: The Cost of Convenience, 2026: Google removed 93 apps after the report (second-hand)
    • Telegram bots: A Large-Scale Study of Telegram Bots, 2026: public data, no private groups
  • crypto
    • Tsuchiya et al., Blockchain Address Poisoning, USENIX Security 2025
      • scammers send dust from look-alike addresses and wait for a copy-paste mistake
      • “270M on-chain attacks targeting 17M victims. 6,633 incidents have caused at least 83.8M USD in losses”
    • Chen et al., From Hype to Collapse, 2026: flags 76,469 of 100,063 new Solana tokens as rug pulls, where the makers sell out and vanish; data released
  • people
    • Oak and Shafiq, “Hello, is this Anna?”, SOUPS 2025: 26 pig-butchering victims
      • victims get hit again by fake “recover your money” services
    • Chen et al. above: of 5,000 wallet-theft victims, “2,019 addresses (40.38%) did not take any remedial measures”
  • I found no academic measurement of
    • fake trading apps
    • fake customer support numbers placed through search ads
    • delivery and refund scams
    • Discord scams
    • no paper that organizes the whole scam field either
    • these are searches that came back empty, so treat them as leads

what works, with evidence

  • make banks pay: the best evidence
    • UK rule since October 2024: the victim’s bank and the receiving bank refund the victim
    • evaluation for the UK payments regulator, read through a law firm’s summary (second-hand)
      • covered losses “fell by around 21%, equivalent to about ÂŁ73m a year”
      • biggest drops at banks that had refunded least before
      • scams paid to accounts abroad rose from ÂŁ21m in 2023 to ÂŁ60m in 2025
        • looks like scammers moved to where the rule does not reach
      • one year of data; slow-to-surface investment scams are left out
  • change what the pay button says
    • Akesson, Gathergood, Quispe-Torreblanca, Preventing Payments Fraud in the FinTech Era, CeDEx discussion paper 2023-08, University of Nottingham
      • 8,958 people in a mock banking app
      • fraud payments fell from 22% to 4% with redesigned buttons plus a risk warning
      • “the effectiveness of behavioural warnings decayed with increasing exposure to fraud”
      • limit: play money, and people were told to watch for fraud
  • kick out accounts
    • the Baidu result above
  • go after the phone companies that let the calls in
    • Prasad, Nahapetyan, Reaves, IEEE S&P 2025, linked below
      • “about 99% of campaigns subject to enforcement actions through the Project Point of No Entry (PPoNE) investigation ceased to operate after the enforcement actions”
        • that was a US FTC action against companies that carry foreign calls into the US
      • those campaigns were 5.5% of the robocalls in the study
  • mixed or missing evidence
    • signed caller ID in the US, called STIR/SHAKEN
      • Prasad, Nahapetyan, Reaves, Characterizing Robocalls with Multiple Vantage Points, IEEE S&P 2025
        • “remaining robocallers have succeeded in in assimilating to the new regime”
        • the paper discusses “why STIR/SHAKEN has failed to dramatically reduce robocall volumes”
    • UK check that the account name matches before you pay
      • regulator’s own words: “some evidence of reduced levels of fraudulent funds”, policy statement PS22/3, no numbers in the lines read
    • Singapore’s registry of text message sender names, Australia’s drop since 2022
      • only government statements found, no outside evaluation
    • freezing stablecoins
      • Wu et al., Ordering Power is Sanctioning Power, arXiv 2026
        • “At least 7.3% of sanctioned USDT addresses and 18.7% of sanctioned USDC addresses had already been drained to zero before the freeze took effect.”
      • Tether’s joint unit froze about $300M by October 2025, against $14B+ of scam inflows in 2025 per Chainalysis (both second-hand)
    • reporting ads to platforms
      • BEUC, Sponsored by Scammers, May 2026
        • consumer groups in 13 countries reported 893 scam ads to Meta, TikTok and Google
        • 243 taken down, 297 rejected, 168 ignored, 185 “removed before review”
        • advocacy study; how each ad was judged a scam is in its annex, which I did not read
    • talking to scammers with chatbots to waste their time or get their bank details
      • Siadati et al., Send to which account?, arXiv 2025
        • 2,638 email exchanges over 5 months
        • where the scammer wrote back at least once, 31.74% gave up payment details
      • Acharya et al., ScamChatBot, arXiv 2024: PayPal confirmed 163 of 743 reported accounts (second-hand)
      • nobody measured whether victims lost less afterwards

AI on both sides

  • scammers
    • Gressel et al., Love, Lies, and Language Models, USENIX Security 2026
      • interviews with 145 people from inside scam operations
      • a week-long test with 22 volunteers, each chatting with one human and one LLM
        • asked to try a harmless app, 46% did it for the LLM and 18% for the human
        • “Our sample (n = 22) was small and drawn primarily from a university population”
      • “popular safety filters detected 0.0% of romance baiting dialogues”
        • makes sense: the early chat is just friendly talk
    • Czybik et al., A Large-Scale Study of Personalized Phishing using Large Language Models, USENIX Security 2026
      • 7,700 people; clicks: LLM personal email 10.0%, generic 3.7 to 4.1%, hand-written personal 24.2%
      • “the cost of personalization is minimal, with approximately $0.03 per email”
    • Heiding et al., Evaluating Large Language Models’ Ability to Automate Spear Phishing, 2026: 54% clicks for AI emails with 101 people (second-hand)
      • 54% vs 10% is far apart
        • the 2 studies differ in people, emails and size, so it may not be a real conflict
    • real use: mostly vendor reports
      • OpenAI: scammers mainly use ChatGPT for “relatively simple tasks like translation” (second-hand)
      • Luu and Samuel, Unintentional Consequences: crypto scam reports rose by “about 722 weekly” after ChatGPT came out
        • reports are not scams, and one date proves little
    • voice cloning: only news stories and marketing numbers found
  • defenders
  • AI agents as the new victims
    • Roy et al., “I Strongly Suspect This Website Is a Scam”, arXiv 2026
      • agents that said the site looked like a scam “still submit critical PII in 35.9% of sessions”
        • PII is personal data such as card numbers
    • several more benchmarks came out in 2025 and 2026: SusBench, WebDecept, TRAP, LoginTrap
      • I think this corner is already crowded

scam ads: loud topic, thin science

  • Reuters, November 2025, on Meta’s internal papers: about 10% of 2024 revenue from scam and banned-goods ads (second-hand)
  • Austrian media regulator, fraud ecosystem study, 2026
    • “Within just three months, the study identified 634,000 fraudulent or problematic advertisements across eight fraud schemes. Together, these ads generated more than 1 billion impressions across the EU”
      • 448,699 of them were online gambling ads
      • found by keyword search in Meta’s ad library
    • “62.4% of all identified ads had already been removed by Meta at the time of data collection”
      • subscription trap ads: 4.5% removed
    • disabled accounts kept advertising: “new ads were later placed under the same Page ID, despite the supposed deactivation of the account”
    • ads vanish from the library: “previously documented ads could no longer be found in later searches”
      • the law says ads stay listed for a year
    • “approximate values based on the data made available by the Meta Ad Library, the accuracy of which cannot be independently verified”
  • Gen Digital (Norton): 30.99% of 14.5M Meta ads in the EU and UK led to a scam, phishing or malware link (second-hand, vendor)
  • EU law gives researchers 2 public data sources
    • each very large platform must keep a searchable library of its ads
    • platforms must log every moderation decision in one EU database
      • Trujillo, Fagni, Cresci, The DSA Transparency Database: “a remarkable fraction of the database data is inconsistent”
      • Kaushal et al., Automated Transparency: platforms have “a lot of discretion” in what they log
      • “scams and fraud” is one of its categories; I found no paper that uses it to study scams
  • I found no peer-reviewed scam ad audit that others can rerun, and none across platforms

research ideas

  1. test the scam protection that already ships
    • question: how many fresh scams do Chrome, Edge, Safari, Android’s message and call detectors, and crypto wallets catch, and how many hours late
    • why it sells: vendors announce AI scam detection with no numbers; an outside scoreboard is easy to explain
    • how
      • take fresh scam sites from a public feed every day, split by kind of scam
      • visit each site in each real browser every few hours and record the first warning
      • for messages: replay the public Reddit scam texts and normal texts into phones
    • first experiment: 500 fresh scam sites of 3 kinds, 3 browsers, 2 weeks
    • measure: share caught at first sight, hours to first warning, warnings on normal sites and messages
    • closest work
      • PhishTime, USENIX Security 2020, did this for phishing
        • it put up fresh test sites and timed each browser’s warning
        • so the method is old; scam sites and phone features are the new part
        • details in the phishing study
      • one-time counts: 0.46% for fake shops, 15.6% for job scam sites
      • Topcuoglu et al. tested LLMs, not shipped products
      • Lu et al. tested off-the-shelf text filters, not the phone’s own
    • stop if: browser detectors only run for logged-in or opted-in users in ways a test rig cannot mimic
    • effort, my guess: medium; phone automation is the hard part
  2. does reporting a scam do anything
    • question: when you report a fresh scam site, wallet, or ad, does it die sooner than one nobody reported
    • why it sells: every victim is told “report it”, and nobody knows which report matters
    • how
      • same daily feed as idea 1
      • report a random half now, the other half after 7 days
      • report to one place per group: registrar, hosting company, Google Safe Browsing, the ad platform, the payment company
      • watch when each site dies or gets a warning page
    • first experiment: 300 fake shops, registrar vs Safe Browsing vs no report
    • measure: days alive after report; for crypto wallets, money received after report, which is public
    • closest work
      • Ctrl+Alt+Deceive: median life 11 days; registrars had taken down 4% of the dead ones
      • random-delay notification tests exist for hacked sites and phishing, as far as I remember
        • Çetin et al., WEIS 2015; Moore and Clayton on phishing takedown
        • not read for this review
      • BEUC: 27% of reported ads removed, no comparison group
      • Evaluating the Effectiveness of Handling Abusive Domain Names by Internet Entities, 2022: could not open it; read before starting
      • PhishTime reported its own test phishing pages and timed the response; see the phishing study
    • snag: a site in the wait group may be reported by someone else
      • check other feeds before assigning, and count it in the analysis
    • stop if: nearly all sites die within a day or two on their own, leaving nothing to speed up
    • needs ethics review: the delayed half stays up longer than it had to
    • effort, my guess: low
  3. scam ads: how many people see one before it comes down
    • question: per platform, how long does a scam ad run, how many people does it reach, and do banned advertisers come back
    • why it sells: the Reuters story, the BEUC complaint, and regulators who need numbers they can check
    • what is new: less than I first thought
      • the Austrian study already has reach, running time and removal share for Meta
      • still open: other platforms, labels from the landing page instead of keywords, daily snapshots so vanished ads are kept, and code others can rerun
    • how
      • pull finance, shopping and crypto ads daily from the EU ad libraries of Meta, Google and TikTok
      • open each ad’s landing page in a real browser from a home internet address
        • scam pages show a clean page to crawlers; the phishing study covers ways around that
      • label a page a scam only with outside proof
        • a regulator warning list, a known scam feed, or a payment account already tied to fraud
        • this misses scams nobody has listed yet, so counts are a floor
      • keep each ad’s first seen date, last seen date, and reach
    • first experiment: 2 weeks, Meta only, investment ads in 2 countries; check 200 labels by hand
    • measure: reach before removal, days to removal, share of scam ads by top 10 advertisers, how fast a removed advertiser’s page reappears under a new name
    • closest work: Austrian regulator study, BEUC, Norton, Ctrl+Alt+Deceive, ScamMagnifier
    • stop if: Google’s and TikTok’s libraries do not mark removed ads or give reach
      • Meta’s does both; I did not check the other two
    • effort, my guess: low for the pilot, medium for 3 platforms

more ideas, weaker or harder

  • time-to-freeze vs time-to-cash-out for scam money in stablecoins
    • freezes are public on the blockchain; Wu et al. and firms like BlockSec already count how much escapes
    • open part: how long after the first public victim report the money is still sitting there
    • snag: the biggest report site refused research access to Gomez et al.
  • how fast does a blocked scammer come back
    • Click Trajectories argued a new bank takes long to get; nobody measured that for merchant IDs, ad accounts, or exchange accounts
    • needs a partner who does the blocking
  • one payment map across scam kinds
    • run victim-acting agents for shop, job, romance and investment scams and record where each asks you to pay
    • pieces exist: Anansi, ScamMagnifier, Siadati et al., ScamChatBot
    • large effort
  • does scam baiting hurt scammers
    • Herley’s theory says wasted scammer time should cut victims; flood one operation and watch whether it answers real people slower
    • hard to observe from outside
  • a payment guard for AI agents
    • stop the agent’s outgoing payment or data when the site is unknown, whatever the agent thinks
    • Roy et al. proposed it and did not build it; crowded area
  • fake support numbers in search ads: no academic measurement found; small, clean audit
  • rerun the two LLM phishing studies, 54% and 10% clicks, with one method
  • use the EU moderation database to compare how platforms act on scams
    • risk: the 2 papers above say the data is inconsistent

limits of this review

  • about 70 sources; roughly 25 read in full text, the rest from abstracts or second-hand
  • “I found no” means a few searches came back empty
  • not covered well: voice cloning, money mule detection inside banks, scam compounds, recovery scams, Apple and Samsung scam features
  • ChatGPT was not consulted: its command line tool needed a login on 7 Oct 2026

8 October closest-work constraint

  • Oest et al., PhishTime, USENIX Security 2020
    • primary abstract: “we systematically launch and report 2,862 new (innocuous) phishing websites”
    • six deployments over nine months measure blacklist speed, coverage, and consistency
    • reading depth in this continuation: official abstract
      • consult the full methods before copying the experimental design
  • distinguish a browser warning from a hosting shutdown
    • reporting that improves warning speed need not remove the site
    • compare warning time, source accessibility, and payment activity as separate outcomes
  • broad browser protection measurement already overlaps this work
    • phone and wallet features need specific closest-work checks

Last edited: