Prompts, shell history, a jailbreak: AI assistants turn up inside a crypto-fraud build pipeline

Rapid7 researchers found an exposed web directory on infrastructure supporting a cryptocurrency fraud operation and pulled down its working files: raw phone-number datasets, account-validation tools, enriched lead records, phishing panels, voice-dialing scripts, fake wallet applications, persistence mechanisms and Telegram exfiltration code ^[1][2]. The part that matters beyond the inventory is the development trail left in the same directory, which turns attacker use of AI coding assistants from an inference about writing style into a set of artifacts.

According to Rapid7, recovered prompts, shell history and project files show AI being used to package Electron applications, obfuscate code, troubleshoot builds, modify phishing infrastructure and prepare malware for distribution ^[3]. That is not snippet generation at the margins; it is the assistant sitting in the build loop. And the loop had friction: when one model began resisting parts of the workflow, the operator switched providers and attempted to bypass the next model's safety controls with a custom jailbreak prompt ^[4]. Rapid7's account does not name the models or providers involved ^[4]. The refusal is worth noting as much as the workaround, because it puts a cost on the operator and leaves a record.

The campaign itself is unglamorous and well-organised. Rapid7 tracks it as Operation ASTERIX, after the Asterisk open-source telephony platform recovered on the server, which the operator used to automate vishing and coordinate calls with phishing emails and counterfeit wallet applications ^[5]. The chain ran from bulk account enumeration against cryptocurrency platforms, through phishing emails that opened fake support cases, into vishing calls that referenced details from those emails, and on to counterfeit Ledger, Trezor and Exodus applications built to steal seed phrases and exfiltrate them over Telegram ^[6].

Scale first. The server held roughly 885,000 phone numbers split into files by region and source, the largest being 316,002 German mobile numbers, with further lists covering Hong Kong and Bulgaria and directories referencing UK, US, Canadian fintech and Ledger-related lists across 54 countries ^[10][11]. The German file alone is about 36 percent of the total ^[16]. Those numbers were then run through validation tooling to find likely crypto holders ^[9]. In one directory, cdc/, Rapid7 found a Go-based tool that submitted numbers to a Crypto.com account-existence endpoint using 300 concurrent threads, retry logic and rotating residential proxies ^[12]. Files named extract_sg_numbers.py and sg_leads_server.py point to additional lead-management and direct-outreach capability ^[13]. Call logs were not recovered, so individual interactions cannot be reconstructed ^[14].

Each stage either narrowed the pool or built trust before the victim was asked to install software or hand over a recovery phrase, which Rapid7 notes gives defenders several points to interrupt the chain before seed phrases move ^[15]. Much of the infrastructure was still live or under development when it was exposed, so Rapid7 Labs was able to notify providers and authorities mid-campaign ^[7]. Disclosure included Apple's security team ^[8].

Watch whether refusal-then-switch becomes a documented pattern in other seizures, and whether the artifact type worth collecting in an incident is now the prompt log rather than the binary.