FraudGPT Underground Chatbot

What Happened

In mid-2023, an actor using the alias "CanadianKingpin12" advertised "FraudGPT," billed as an unrestricted, ChatGPT-like AI chatbot for cybercriminals, on multiple dark-web marketplaces (Empire, WHM, Torrez, World, AlphaBay, Versus) and, after forum threads were repeatedly taken down, primarily via a Telegram channel created 2023-06-23. Netenrich's threat research team documented the listings publicly on 2023-07-25, noting the tool had circulated on Telegram since 2023-07-22, and described FraudGPT as marketed for writing spear-phishing/BEC emails, malicious code, "undetectable" malware, phishing pages and panels, scam letters, and for locating non-VBV bins, cardable sites, leaks and vulnerabilities, with 24/7 escrow support. Subscriptions were priced at $200/month or $1,700/year, and the actor claimed over 3,000 confirmed sales. Netenrich explicitly framed FraudGPT as similar in kind to WormGPT (reported by SlashNext on 2023-07-13), and traced the operator to the email canadiankingpin12@gmail.com. SlashNext separately engaged the actor undercover and reported claims of two additional in-development bots, "DarkBART" and "DarkBERT," with internet access and image capabilities. Neither Netenrich nor SlashNext reported any confirmed real-world attack tied to FraudGPT-generated content at the time. The story took a significant turn in 2025, when Cisco Talos published research describing its own attempt to purchase FraudGPT access: after extended negotiation with CanadianKingpin12 on Telegram, Talos received a username and password that did not work, and the actor then demanded cryptocurrency for a "crack" to the login page, a pattern Talos assessed, and multiple other scammed buyers corroborated, as proof the actor had no working AI product at all and was running an advance-fee cryptocurrency scam under the FraudGPT brand.

How the Attack Worked

CanadianKingpin12 first tried to sell FraudGPT on lower-level clearnet cybercrime forums; threads were repeatedly removed, so the actor pivoted to a Telegram channel (created 2023-06-23) to distribute more resiliently, avoiding the exit-scam risk of dark-web markets like Empire, WHM, Torrez, World, AlphaBay and Versus where the actor claimed "verified vendor" status. The Telegram/dark-web listings advertised a ChatGPT-style interface with no ethical guardrails, promoted with a features list (malicious code, undetectable malware, phishing pages/panels, scam letters, non-VBV bin lookup, cardable-site discovery, vulnerability/leak search, page hosting, code obfuscation, OTP-spoofing bots, CVV checking, and 24/7 escrow) and a claim of 3,000+ prior sales. Netenrich's threat research team discovered and documented the listings on 2023-07-25, tracing the operator to the email canadiankingpin12@gmail.com. In 2025, Cisco Talos independently engaged CanadianKingpin12 on Telegram posing as a buyer; after prolonged negotiation the actor supplied a username/password for the FraudGPT site that did not work, then demanded cryptocurrency to purchase a "crack" for the login page, behavior Talos assessed, and other scammed buyers corroborated, as evidence the actor had no working AI product and was running a straightforward advance-fee/crypto scam under the FraudGPT brand.

The Lure & the Tell

Lure (to prospective cybercriminal buyers): a slick "exclusive bot" pitch positioning FraudGPT as a ChatGPT alternative "with no boundaries," backed by a long feature list and a claimed track record of 3,000+ sales/reviews: social proof and scarcity framing (limited licenses, per-month virtual-machine/page-hosting caps) designed to create urgency and legitimacy for buyers on cybercrime forums. Tell/giveaway: when Cisco Talos actually attempted to purchase access, the promised login credentials simply did not work, and the actor pivoted immediately to demanding a further cryptocurrency payment for an unexplained "crack," a classic advance-fee-fraud pattern (moving the goalposts after initial payment) that other buyers on the same forums independently reported experiencing.

Outcome

Netenrich's 2023-07-25 report made FraudGPT one of the most widely cited examples (alongside WormGPT) of the "malicious LLM-as-a-service" trend, driving broad vendor and media coverage (SC Media, SecureWorld, Security Boulevard, Varonis, etc.) through mid-to-late 2023, and prompted general law-enforcement commentary: FBI Director Christopher Wray publicly warned on 2023-07-26 that criminals were exploiting generative AI for malware and spearphishing, without naming FraudGPT specifically. Netenrich stated as of its report date it knew of no confirmed active attacks carried out using FraudGPT-generated content. No public record of a specific victim organization or law-enforcement case tied to FraudGPT-produced phishing was identified. In 2025, Cisco Talos's own attempted purchase established that the offering was very likely a scam targeting cybercriminal buyers rather than a functioning AI weapon, undercutting the tool's own advertised capabilities, though the "FraudGPT" brand persisted in underground and vendor-report circulation into 2026 as a loosely-used label for copycat/rebranded "uncensored AI" offerings.

Why It Matters

FraudGPT is one of the earliest and most-cited "malicious LLM-as-a-service" offerings, widely referenced in threat-intel literature (alongside WormGPT, GhostGPT, DarkGPT, DarkestGPT and others) as evidence that generative AI lowers the technical barrier to phishing and fraud. But the case is equally important as a caution against taking underground-market capability claims at face value: years of follow-on vendor research (Cisco Talos in 2025, echoed by Rapid7's 2026 criminal-AI-market analysis) found no verified working FraudGPT backend, and concluded the operator was very likely running a straightforward scam against fellow criminals rather than operating a functioning offensive AI tool. For an educational audience, the incident illustrates two overlapping social-engineering layers: (1) the advertised threat, AI marketed to make phishing/BEC content more convincing and scalable, and (2) the demonstrated threat, that the "AI cybercrime tool" market itself is rife with advance-fee scams that use the same urgency, social-proof, and authority tactics found in consumer-facing fraud, just aimed at criminal buyers instead. It's a useful corrective against uncritically amplifying vendor claims about "new AI attack tools" without independent verification.

Defenses

Treat "blackhat AI" marketplace claims (FraudGPT, WormGPT, DarkGPT, GhostGPT, etc.) with the same skepticism as any other unverified cybercrime-forum product. Vendor threat intel repeatedly finds these are frequently non-functional or vastly oversold, so downstream defense should focus on the actual delivery mechanism (phishing emails/pages, BEC lures) rather than the AI branding. Standard anti-phishing/anti-BEC controls remain the effective countermeasure regardless of whether AI generated the lure: DMARC/SPF/DKIM enforcement, execution-time link/attachment sandboxing, out-of-band verification for payment/wire-change requests, security awareness training that flags urgency+financial-request combinations, and behavioral/email-gateway detection tuned to the writing-quality and infrastructure patterns of BEC rather than to "AI-generated" as a distinct signature (AI-polished text is often indistinguishable from human-written text). For organizations evaluating cyber-threat-intel vendor claims about new "AI crime tools," corroborate with a second independent source (as Talos/SlashNext/Netenrich did here) before treating capability claims as fact.