Anthropic published a report on how its models were used in attempts at hacking, surveillance, and propaganda: nine months of observations from December 2025 to August 2026, covering thirty-seven flagged operations. Amidst headlines about AI hackers, a detail crucial for the Telegram audience is being overlooked: the word Telegram appears fifty-three times in the report, while "cryptocurrency" appears not once.
No "Bitcoin," no "wallet," no "blockchain." The only instance of "crypto" in two hundred sixty thousand characters is "cryptographic software" in a table of Russian defense procurements, referring to cryptography, not coins. Meanwhile, Telegram is mentioned more often than WhatsApp, X/Twitter, and Instagram combined.
How Telegram Was Used in These Operations
Not as a messenger for attackers to communicate. In the report, it performs six different roles, almost none of which are its usual descriptions.
| Role | Description |
|---|---|
| Conveyor Belt | Verified stolen keys drop into a group organized by topics |
| Exfiltration Channel | “Telegram bot streams” is one of six ways data is exfiltrated from victims |
| Warehouse | A group chat where loot is stored and traded |
| Storefront | A stolen card shop built as a Mini App |
| Control | Malware C2 channel, with chat IDs listed in indicators |
| Disguise | An executable file named Telegram with its icon |
And a seventh role — target: private group data dumps, participant deanonymization, and a mass reporting bot are described in the surveillance section.
Conveyor Belt Organized by Chat Topics
An operator using the aliases MeowSHA, frkoo, and blazespider — an affiliate of the ShinyHunters group — maintained ten worker machines on AWS EC2. They downloaded 1.8 million different Android APKs from several app stores, decompiled them, and searched for embedded secrets using the TruffleHog scanner. The findings were checked for validity and streamed in real-time to a Telegram group, categorized into over a hundred source types.
This is where the mechanics worth reading the indicator tables begin. There are two groups. The first wave of verified keys went into "ClintonHog." The second, "ChatMignon," contained 471 forum topics, one for each type of secret detector. Forum topics in a supergroup are designed as discussion categories; here, they became a routing table: an AWS key drops into its branch, a GitHub token into its own, a payment gateway key into its own. TruffleHog has over eight hundred detectors, so roughly half the catalog yielded at least one confirmed hit and earned a dedicated topic.
Sorting stolen goods is the most tedious and labor-intensive part of a fence's job. Here, it wasn't coded; it was adopted ready-made from the messenger.
Stolen Card Shop as a Mini App
The operator's storefront was hosted on a domain mimicking the French national police — policenationale[.]cc; Anthropic specifically notes that this was not a phishing lure but branding for a criminal shop. The merchandise — "fiches," payment card records enriched with BIN lookups, full cardholder details, and an interactive map of victim addresses.
All this was delivered to the buyer via a Telegram Mini App, backed by the Soraki platform using PostgreSQL and GraphQL. The same platform also held consolidated French data breaches, including a four hundred thousand-record array from a telecom operator with IBAN and BIC, and made them searchable.
The stack is described in the exact same terms used by legitimate Mini App studios: bot, Mini App, GraphQL, Postgres. Only the product differs.
Where Crypto Appears in the Report
It's not a theme, but it is a target, appearing in three different scenarios.
Fake domains. In the indicators of compromise, alongside email lures, are ari-chain[.]com, arichain[.]network, bitmart-mystery[.]com, and defi-claim[.]xyz. The first two mimic Arichain — a real L1 with an audience awaiting an airdrop. The third mimics a mechanism that the BitMart exchange actually conducts: a "mystery box" with a reward. The fourth doesn't pretend to be anything specific: a claim page as a genre.
Account factory. Another actor, tracked in the report as GTG-50020, industrialized account registration: residential proxies, anti-detect browsers, bots completing exchange and marketplace registration forms, commercial CAPTCHA solvers, automated email polling, and automated identity verification steps. Ready-to-use verified accounts were stockpiled.
Verification interception. For the same actor, victims were redirected to a spoofed verification page whose reverse proxy relayed the genuine KYC flow. The person underwent authentic verification — with real documents — while the operator intercepted the session and documents mid-process, then logged in from their own machine.
Money is plentiful in the report, just not in coins: cards, extortion ranging from one and a half to two and a half million dollars for 26 gigabytes of exfiltrated data, reselling access, and stolen keys from AI providers. One operator, by their own admission, also received legitimate bug bounty payouts of two thousand and five thousand dollars from companies they themselves had hacked and blackmailed.
The word laundering appears fifteen times in the report and never refers to money: it's about laundering the source, attribution, and confidence — to make a state narrative appear as an independent voice. TON is not mentioned at all.
What This Means for You
Verification interception breaks the usual "is this a real site" check. The page is real, the flow is real, documents go to a real service — only the path you took to get there is faked. Only one thing works here: do not start verification from a link in an email, chat, or search; only initiate it from your own login to the service.
A Mini App within Telegram is a web page in a wrapper, not a verified application. A stolen card shop was delivered exactly this way: the platform doesn't distinguish between a fence's storefront and a game.
If you publish a mobile application, assume its APK has already been downloaded and decompiled. One million eight hundred thousand files on ten machines can be analyzed in weeks, and an embedded bot token or payment gateway key can be found by a scanner in seconds.
A verified exchange account is no longer proof of a human behind it: factories produce these on an assembly line, complete with KYC completion.
What Account Numbers Reveal
At the end of the chapter on the operator, there's a table with the Telegram addresses of their infrastructure: the individual, four of their bots, and two groups. An address here is not a username like @durov, but an internal number Telegram assigns upon registration.
The numbers are sequential, so the table dates itself without a single query to Telegram. The operator's personal account is noticeably older than their bots: they used the messenger long before setting up their conveyor belt. The bots, conversely, are almost consecutive — they were created in a single setup session for a specific task. Both groups fall within the range where Telegram places recently created rooms.
This means the operation, which appears well-oiled in the report, was assembled shortly before it was shut down.
How Telegram Was Used in Other Chapters
In the surveillance section, the model translated chatter, bulk-extracted from over a hundred WhatsApp groups and dozens of Telegram channels, into structured Chinese data. Also mentioned were a custom-made messenger deanonymizer, a "phone number → identity" resolver, and a mass reporting bot. Analytics on a fresh dump of a private group with 244 members, including an analysis of internal connections, were separately noted.
In malware, Telegram serves as both cover and conduit: a keylogger recorded keystrokes while Telegram Desktop was in focus; a screenshot component was distributed as an executable file named Telegram with its icon, and screenshots were exfiltrated via a bot; a fake ESET NOD32 antivirus login page sent captured passwords to the same destination.
In propaganda operations, channels became both raw material and platform: source material for Sputnik articles was taken from Telegram channels, and the results were published there. The most disturbing episode in this part was an operation where an agent was tasked with cloning a real activist's personal account: the model read about 8,400 of their posts, copied their style, and engaged in live political conversations with their contacts.
