The State of Scams in Digital Communities — Telegram Edition, 2026

1. The scale of it

Four figures set the scale of community-targeted fraud in 2025. They measure different things, so they do not add up to one total, but together they show an industry, not a nuisance.

The FBI's Internet Crime Complaint Center counted $8.6 billion in reported investment-fraud losses in the United States for 2025, the largest category it tracks and roughly 40% of all reported internet-crime losses. Within it, $7.2 billion was crypto-investment fraud: the long con in which an attacker builds a relationship over weeks or months, usually starting on a dating app or in social-media DMs and often moving to Telegram, then introduces a fake investment and drains the victim. Romance fraud proper adds about another $900 million, counted separately. [1]

Telegram banned 43.5 million groups and channels in 2025. Before the arrest of its founder in August 2024, daily takedowns ran at 10,000 to 30,000; afterwards they rose to 80,000 to 140,000 a day, with peak days above 500,000. [2]

Between November 2024 and January 2025, malicious Telegram-group scams, fake verification bots and malware-laced "trading groups", rose by 2,000%. Part of that is the platform clean-up churning everything; most of it is attackers scaling on AI tooling that had become cheap enough to use at volume. [3]

And the largest illicit online marketplaces in history were not on the dark web. They were Telegram channels. Huione Guarantee and Xinbi Guarantee, Chinese-language "escrow" markets for scam tooling, laundering and the brokering of trafficked workers, processed more than $35 billion in stablecoins before Telegram banned them in May 2025. [18] What happened after that ban is in section 5.

Drawn to scale

Hearing "billions" four times flattens the numbers. Side by side, the largest bar is the industry itself, Southeast Asian scam-centre revenue at about $40 billion a year [21]; the marketplaces at $35 billion in cumulative flows are its supply chain; the $8.6 billion in reported US losses is only what victims told the FBI; and North Korea's $2.02 billion in 2025 crypto theft, the headline-grabber of the year, is the smallest bar on the chart. [23] A state-sponsored hacking programme is a rounding error next to the scam-compound economy.

Why now

Three forces converged. Telegram is the de facto social layer of Web3: every token launch, airdrop and project community runs there, so the money moved there and the attackers followed. Attack tooling became a commodity and then collapsed in price: jailbroken LLM products are sold on Telegram itself, from about $200 a month in 2023 to $220 for a lifetime licence in 2025 (section 4). And the moderation most communities run was built for the pre-AI threat: keyword lists, signatures, simple rate limits, none of which see a conversational, personalised, multi-turn scam coming.

The result is that communities are dying in two ways: to the spam itself, because the tools cannot keep up, or to the over-moderation admins deploy in desperation, which drives real members away.

2. Three eras

Scams on Telegram arrived in three eras. Each raised the stakes on the last, and none of them retired.

Era I, pre-crypto: noise. Scripts, mass invites, keyword spam, drive-by promotion. The same payload thousands of times: hash it, block it, done. A nuisance rather than predation, because there was little in the room worth stealing. This is what classical rule-based moderation was built for, and it handles it well.

Era II, crypto, pre-LLM: fraud. Money entered the room and spam became fraud: pig-butchering, fake-admin DMs, wallet drainers, fake giveaways. The stakes jumped from wasted attention to drained wallets. But the attacks were still run by people, so they stayed labour-bound and left seams: the broken grammar, the copy-paste cadence, the reply from the wrong time zone. This is where classical tools started losing.

Era III, LLM: AI-scaled. AI writes the content and runs the personas. The labour bottleneck is gone: one operator runs thousands of simultaneous conversations, every message unique and on topic, with no signature to match. This is the youngest era and its sophistication curve is still bending upward.

Today's attacker stacks all three: legacy spam for cheap volume, crypto-era playbooks, and LLM tooling to run them at scale. A defence has to hold across all three at once. One thing this framing leaves out on purpose: the problem of one actor behind many accounts is not an era. It is the structural thread under all three, and it has its own section.

3. Ten ways modern attacks defeat classical tools

These are the concrete patterns that walk past rule-based moderation, roughly in order of prevalence. The first three get the most detail.

01 · The clean-chat admin-impersonation DM

The dominant scam vector of 2025 by victim losses. A scammer joins a public group and watches. When a member asks a question or voices frustration, the member gets a private message within minutes, from a profile whose display name and avatar copy an admin's, often with a one-character substitution (a zero for an O, a lowercase L for a capital I). The message in the group, if there is one at all, is benign. The payload, a seed-phrase request or a fake KYC link, lives only in the DM, where no moderation bot can see it. [4] [5] [6]

Classical tools miss it for three reasons: the public message has no banned word, link or media; nothing compares a new member's name and avatar against the real admins at join time; and nothing tracks "this account joined three minutes ago and is already messaging active members". The defence is profile-similarity checks at join, behavioural signals, and a report path for DMs the bot cannot observe directly.

02 · LLM-generated conversational spam

The slow pivot. An account joins and posts on-topic, natural messages that match the group's culture. It answers questions, correctly. It builds plausibility over several turns, sometimes days. Then it pivots: a link, an offer, a request to continue in DMs. Every message is generated, not copied, so there is no duplicate to detect, no signature, and the tone matches because the model was told to match it. The tooling is commercial: FraudGPT listed at about $200 a month in 2023 [7]; WormGPT 4 sells for $50 a month or $220 for life, source code included. [8] Nothing that looks at a single message catches intent that only appears across a conversation; the detection that does has to read the rolling window, and in practice that means another model.

03 · The edit-after-post payload swap

The simplest bypass in this report, and most bots are still open to it. The spammer posts a clean message that passes every filter, waits out the new-joiner restriction window, then edits the message to insert the link or the scam text. Most moderation bots subscribe to Telegram's message update and not to edited_message, so the edit never enters the detection pipeline. The fix is a day's work: process both events and re-run every check on an edit. If you run a community, ask whoever maintains your bot whether it processes edits.

04–06 · The signature problem

Three patterns with one root cause: classical detection matches signatures, and modern attacks break the surface form.

• Homoglyph and zero-width obfuscation. A Cyrillic а, е or о in place of the Latin letter, or an invisible U+200B/U+200C/U+200D inserted mid-word, splits "airdrop" or "metamask" so a blocklist no longer matches while the reader sees nothing wrong. Unicode normalisation and a confusables map before any keyword check fixes it; most bots do not do it.
• Image and screenshot scams. The URL, wallet address or "official" handle is rendered inside a JPEG, a sticker or a screenshot of a fake pinned message. Without OCR the text is invisible to the bot, and the usual answer, blocking all media, is unusable in a living chat.
• Aged-account swarms. An industry farms Telegram accounts three months to five years old, with plausible history and avatars, and sells them wholesale specifically to pass account-age checks. [14] No single-account check tells them apart; only coordinated-behaviour detection across accounts does.

07–10 · The behavioural problem

Four patterns where the attack is not in what was said but in how the account behaved or what context it brought.

• CAPTCHA bypass and CAPTCHA fatigue. Solver services defeat challenges for pennies, while challenges strong enough to slow attackers also cost real members, who give up and leave. The answer is risk-adaptive challenges that only the suspicious joiner sees, which few deployed bots do well.
• The lurker-to-DM funnel. The account joins, never posts, scrapes the member list and opens DMs. It trips no in-group rule, and it is the mechanism behind a large share of the long-con losses in section 1. [1]
• Hyper-personalised phishing. Breached data plus a language model produces a message that references the victim's own project, wallet or on-chain activity. It reads as real outreach. The scam is in the specificity, which has no signature.
• Malicious "support" bots. Bots named for major wallets and exchanges. Third-party verification for accounts and bots has existed since January 2025 [32], but adoption is minimal and any bot can still launch under any name. [4]

4. The AI threat in detail

The economics

When FraudGPT surfaced in July 2023 it was sold on a subscription, from $200 a month, advertised on Telegram with thousands of claimed sales. [7] By November 2025, WormGPT 4 sold for $50 a month or $220 for a lifetime licence with source code. [8] A year of the tooling went from about $2,400 to $600, or $220 once: a 91% fall in two years, and newer entrants undercut even that. When attack tooling deflates that fast, attacker volume is the output.

The distribution channel is Telegram itself. The platform being defended is the marketplace for the tools used against it: a closed loop.

The clearest large-scale before-and-after measurement is SlashNext's finding of a 1,265% rise in phishing volume in the twelve months after ChatGPT launched (Q4 2022 to Q3 2023). [9] That is a correlation with the generative-AI era rather than a strict attribution, but nothing since suggests the curve bent back down.

What the tooling can do: hold a coherent multi-turn conversation, match the register of a given channel, and personalise from breached data and public profiles.

Why signatures fail

Before AI, the same scam message was pasted to thousands of victims; hashing it once blocked every repeat. That is why moderation bots evolved around keyword matching, regular expressions and blocklists. After AI, the same intent ("send me your seed phrase") appears in endless surface variations, so the hashes never collide and the keywords never match. Detection has to move from what was said to what the actor is doing: how the account joined, what it did before posting, how it engaged, when it pivoted, what identity and reputation it carries. Judging intent over a conversation window with another model works, and is expensive; cost discipline is one of the open problems of the field (section 8).

The agentic shift

Two 2026 cases show where the AI threat now sits: not generating spam, but running the campaign.

In May 2026, Trend Micro published its investigation of "Patriot Bait". One Russian-speaking operator ran a 17,000-subscriber Telegram channel impersonating an American veteran. The channel had existed since February 2021; from September 2025 to May 2026 it ran on AI. A Python orchestrator the operator called "Quantum Patriot" called a frontier model through 73 stolen API keys to play the persona, and a separate chatbot talked to victims directly. The operator jailbroke the model by convincing it he was an authorised penetration tester, a role the model saved to memory, so every later conversation started from the jailbroken state. In Trend Micro's words, what previously required a team of writers, social-media managers, IT workers and malware programmers can now be automated by a single actor with a server, a Telegram bot and API access to frontier models. [10] The ScamAgents paper formalises the pattern: autonomous multi-turn agents with memory and adaptive behaviour, against which current safety guardrails are documented as ineffective. [11]

For detection this cuts both ways. Classical tooling assumed scam-team economics, many distinct fingerprints per campaign. A solo agent collapses that: one IP, one model account, one operational style, which is easier to fingerprint at the campaign level and harder at the message level, because every message reads like one thoughtful person.

The Truman Show

In January 2026 Check Point published its investigation of a campaign it had tracked since October 2025, built around an app called OPCOPRO and named for the 1998 film about a man whose whole world is a set. Tightly controlled Telegram and WhatsApp groups, often synchronised, are populated with AI-generated "experts" and synthetic peers. Members ask questions, peers answer, experts advise. The immersion runs for weeks. Companion apps in the Apple and Google stores are WebView shells whose balances, trades and agreements are fabricated server-side, so victims watch their gains grow. Then comes the ask: invest more, pay the withdrawal fee, cover the tax. [12]

This defeats everything above it in this report. There is no spam message; the entire context is the attack. Every per-message detector misses, every lurker-to-DM heuristic misses, because the community itself is fabricated. What catches it is cross-platform reputation, recognising the same fingerprint across many groups, or identity verification strong enough to unmask synthetic members. The threat model has moved up a layer: from attacks inside real communities to communities that are the attack.

5. One actor, many identities

The other hard problem alongside AI, and the structural one. In security it is called the Sybil problem: one actor controlling many accounts to manipulate a system.

The economics

Aged Telegram accounts sell for $1 to $10 each on adversary marketplaces, three-month to five-year-old accounts available in bulk. The pricing comes from the farming vendors' own marketing, which publishes step-by-step farming guides as sales collateral; the industry is normalised enough to advertise. [14] A medium ring runs 50 to 200 accounts in one coordinated operation; professional operations reach into the thousands, often driven over Telegram's MTProto protocol. Solidus Labs documented one ring, PumpCell, running a synchronised multi-chain pump-and-dump across dozens of Telegram channels that netted about $800,000 in a month. [15] On marketplace prices a 100-account swarm costs on the order of $500; a single successful scam pays $5,000 to several million. The ratio, not the exact figures, is the point.

Why it breaks defence

Every classical defence shares one assumption: each account is an independent actor. Coordinated accounts break it. Bans do not matter when the actor has 50 spares and can buy another for a dollar. Account-age checks do not help when aged accounts are a product. Reputation resets when the swarm rotates. Federated banlists, among the best classical defences, fight a hydra. The structural gap is that no deployed moderation tool can collapse N accounts back to one actor. Without that, defence is whack-a-mole with unlimited moles.

The hydra at platform scale

The proof at the largest scale that exists. Huione and Xinbi Guarantee had processed more than $35 billion in stablecoins, mostly USDT, selling scam tooling, laundering, SIM banks, aged accounts and drainer kits, and brokering trafficked workers for scam compounds. [18] On 13 May 2025 Telegram carried out its largest targeted takedown, removing thousands of associated channels, after Elliptic's research and the US Treasury's designation of Huione Group as a primary money-laundering concern. Within weeks, Tudou Guarantee, a competitor in which Huione had bought a 30% stake five months before the ban, saw roughly a 70× rise in daily volume; Xinbi returned under the same account ID with inflows up 90%. [19] [20] The takedown worked exactly as designed. The market did not slow down, because the actors behind it were never collapsed to identities that could be excluded.

6. Beyond the channel

The biggest losses of 2025 did not end on Telegram. They started there.

North Korean state-affiliated groups stole about $2.02 billion in crypto in 2025, up 51% on the year, with more than $300 million attributed to deepfake Zoom and Teams meetings in which attackers posed as trusted industry figures. [23] [24] One case shows the shape: THORChain's co-founder lost about $1.35 million after a message from a friend's compromised Telegram account led to a video call with a deepfaked contact and then to macOS malware that drained his wallet. [25] Voice-note deepfake distress scams cost about $5 million in 2025, and three seconds of source audio is now enough for consumer voice-cloning services to produce a convincing clone; a single voice note posted in a group chat will do. [26]

The chain has four steps. Entry: a benign DM from a compromised friend's account ("hop on a quick call?"), with no link, no keyword and no new account to flag. Pivot: a calendar invite or look-alike meeting domain, one click off the platform. Deepfake: a live call with cloned faces and voices, then "your camera isn't working, install this fix". Drain: the fix is malware. The only step a Telegram defence can observe is the first, and the first has nothing to detect. Cross-platform reputation and out-of-band verification habits are the only defences that even apply, and no tool shipped in 2024–2026 addresses the cross-platform funnel.

Who runs the scams

The supply side is industrial, and increasingly indicted. UNODC estimates Southeast Asian scam centres generate about $40 billion a year, and names Telegram as convenient and scalable for illicit actors. [21] Those centres run substantially on forced labour: roughly 100,000 people trafficked into compounds in Cambodia and 120,000 in Myanmar, by the same estimates. The "scammer" messaging a community is, in a large share of cases, a trafficking victim working off a fabricated debt. In October 2025 the US Department of Justice indicted the chairman of Cambodia's Prince Group for operating forced-labour scam compounds and seized 127,271 BTC, about $15 billion, the largest forfeiture in US history. [22] The same month, a raid on the KK Park compound in Myanmar freed more than 2,000 workers from one compound among hundreds. For a community operator this sets expectations: the adversary is an industry with the revenue of a mid-sized state, and the moderation bot is the first line.

7. Where the money is

Crypto communities carry a disproportionate load for four reasons. Value per victim: a successful long con nets six or seven figures, an order of magnitude above consumer phishing. Attack surface: crypto groups are open by design, and every member is one tap from connecting a wallet. Documented campaigns: PumpCell is one ring of many. [15] Brand impersonation: fake "support" bots and channels for major wallets and exchanges remain endemic while verification goes unadopted. [4] [5]

Different ecosystems take different attacks. On TON, Telegram-native distribution makes Mini App phishing the dominant pattern, with wallet drains through the in-app webview. [5] On EVM chains, Ethereum and Base take the volume in wallet drainers, fake airdrops and malicious signing prompts that grant token approvals disguised as claims. On Solana, memecoin pump rings, fake launches and drainer bots that automate seed-phrase phishing. Cross-chain, PumpCell-style actors run campaigns on several chains at once and launder across bridges; single-chain operations are old news.

TON in the crosshairs

The Telegram-native trust UI is the attack surface. Fake Fragment clones imitate the username-auction market and drain wallets through TON Connect when users "claim" or "list". "TON Giveaway" and "Wallet Support" Mini Apps request wallet authorisation for fake events and fake support flows. And malicious contract approvals via TON Connect look like a routine sign-in while granting drain permissions. [27] [28] Mini Apps run inside Telegram, trust is implicit, and no moderation tool scans Mini App manifests, contract approvals or Mini App reputation. The same flow defenders would like to use for identity verification is the one attackers exploit today; any defensive work on identity inherits this surface from day one.

The drainer economy

Wallet drainers are sold, supported and updated as software, on Telegram. Inferno Drainer "retired" in 2023 and returned in 2025 with command-and-control addresses stored encrypted on-chain, so there is no server to seize, and proxy infrastructure run by its own customers; Check Point documented about $9 million drained in its first half-year back. [29] When Inferno's operators stepped back they handed the codebase to Angel Drainer, whose AngelX has deployed more than 300 malicious dApps targeting TON and Tron specifically. [30] TrickMo, an Android banking trojan, moved its command-and-control onto TON's own peer-to-peer overlay in 2026: no DNS, no hosting provider, nothing to take down. [31] This belongs in a report about community scams because every drainer needs a delivery channel, and the delivery channel is the community: the fake airdrop in the group, the Mini App link in the DM.

8. What still works, what has plateaued

What still works

Six things still earn their place, and the new layers belong on top of them, not instead of them.

• Join-time CAPTCHA with restrict-on-join. Vendor roundups report around 95% spam reduction; treat the figure as directional, not audited, but it matches operator experience as the single highest-return filter. [16] The answer to CAPTCHA fatigue is risk-adaptive challenges, not removing them.
• Locks on URLs, forwards, media and invites kill the long tail of low-effort spam at almost no cost.
• Federated banlists such as CAS propagate one community's ban to every subscriber: cheap, precise, free wins. [17]
• Anti-flood rate limits still defeat the dumb half of swarms.
• Member reports are cheap human-in-the-loop signal that most communities underuse.
• Account-age checks catch casual attackers, though not professional farms.

What has plateaued

Rule-based detection against AI-generated content, because there is no signature when every message is unique. Account-age heuristics against farmed aged accounts, because the farming industry commoditised the defence. Single-platform reputation, because attackers cross platforms and bans do not. Static keyword filters, because homoglyph and zero-width tricks defeat them silently, so the bot reports clean while the bypass walks through. None of these is worthless; they still catch the casual end of the distribution. They have stopped being sufficient.

The platform's own response

Telegram moved more on abuse in eighteen months than in the prior decade, and an honest account has to say so. In September 2024 it reversed its data-sharing stance: IP addresses and phone numbers now go to law enforcement on valid request, and fulfilled US requests went from 14 a year to about 900. [34] In January 2025 it launched third-party verification for accounts and bots, the anti-impersonation primitive defenders had asked for; adoption so far is minimal. [32] In 2025 it introduced paid messages, per-message fees for non-contacts, economic friction aimed at the DM funnel. [33] And it carried out the mass takedowns of section 1. Attackers adapted: pre-staged backup channels, "reserve" handles announced in advance, and request-to-join gating that also locks out automated moderation. Migration off Telegram barely registers; the audience is on Telegram, so the criminals stay. [2]

Enforcement throughput rose about seventeen-fold, from up to 30,000 takedowns a day to up to 140,000, with peaks above half a million. Every attack pattern in this report operates at full scale on top of that curve. Takedown volume is necessary and visibly not sufficient; the remaining gaps are structural and sit at the community level.

9. What is emerging

Three directions, with the "how" still unsolved in public.

AI-assisted detection. Using models in the moderation pipeline is happening because it has to; signature detection is dead against generated spam. The open problem is cost: a model on every message blows up the bill at any real volume, so the architectures that work route carefully and reserve the expensive judgement for the messages that need it.

Behavioural signal layers. Detection that looks past what a user said to how the account behaved: how it joined, what it did before posting, how it engaged, when it pivoted. Mostly absent from deployed bots, and the largest missed opportunity in the field.

Verifiable identity. Wallet-, credential- or proof-of-human-gated trust for high-stakes contexts. It is the only thing that breaks the one-actor-many-accounts economics, because an identity anchored to something with cost cannot be bought 200 at a time for $500. Adoption is the bottleneck; for DAO voting, alpha channels and paid memberships it is a defensible direction, one piece of a converging answer rather than the whole of it.

Choosing among these, and operating any of them at scale, is the work of the next two to three years. Predictions and an operator playbook follow in the second part of this series.

Methodology

This report is written from the May 2026 talk of the same name. Every figure was re-checked against its primary or first-reporting source before publication, and each carries a numbered marker to the list below. Where a figure comes from vendor marketing or a vendor roundup, the text says so. Illustrative arithmetic (the cost of a swarm against the payout of a scam) is labelled as such. Nothing here states a conclusion about a named company or product beyond what its cited source states. Corrections are recorded in the changelog and bump the version.

Targa-Praetorian (2026). The State of Scams in Digital Communities — Telegram Edition, 2026 (v1.0). https://targa.dev/research/state-of-scams-2026

@techreport{targa_state_of_scams_2026,
  title  = {The State of Scams in Digital Communities — Telegram Edition, 2026},
  author = {{Targa-Praetorian}},
  year   = {2026},
  url    = {https://targa.dev/research/state-of-scams-2026},
  note   = {v1.0, CC BY 4.0}
}