AI-Assisted Cybercrime: How One Operator Used Hermes and DeepSeek Across Attacks

One operator, several attack paths

AI gave one cybercriminal a way to work across attack types that usually call for different skills. Unit 42’s investigation found an operator using AI for reconnaissance and code, while also seeking detailed advice on VoIP fraud. The activity targeted multiple networks, primarily in India, and included a confirmed intrusion that exposed personal information on more than 5,200 students.

The operator, known as BUGSTER, configured the Hermes Agent framework with DeepSeek V4 as its reasoning engine and used Cursor as a secondary coding tool. Both were set up with safety controls disabled. Hermes provided the main interface through Telegram direct messages, with specialized profiles running as separate bot gateways. The actor also configured profiles named dev, check, vuln and deploy in a shared group. Investigators found no chat records for those profiles, though their names suggest roles in development, validation, vulnerability testing and deployment.

The setup points to AI being used as an operating layer across several kinds of work, rather than for one isolated task. The actor generated scripts and structured reconnaissance reports, and used different bot profiles to seek advice in particular areas. That breadth helped a single operator attempt activities that would ordinarily require separate technical knowledge.

University intrusions and exposed data

At one Indian university, AI-assisted reconnaissance supported a methodical attack on an online education platform. Unit 42 described a four-stage progression, from mapping the target through to brute-force activity. The actor’s WordPress REST API enumeration report used executive summaries, severity ratings and organized endpoint categories—formatting investigators considered consistent with AI-generated output. A brute-force attempt recorded a candidate credential.

The operator also broke into a second Indian university and extracted a MongoDB database containing more than 5,200 records, including students’ personally identifiable information. The extraction tool created a manifest alongside the dumped database collections. The source report does not specify the contents of that manifest, so its role is clearest as a record of what the tooling collected.

The actor also enumerated a site on a popular e-commerce platform and extracted information from it. Unit 42 describes the activity but does not provide further details about the data or any resulting use. Separately, the actor’s suspected AI-generated attack report included a section titled “NEXT EXPLOIT OPPORTUNITIES,” laying out possible follow-up attacks. That document suggests the operator was using AI to turn reconnaissance into a queue of prospective next steps.

VoIP fraud guidance, still at an early stage

One Hermes profile, alice, served as an AI adviser on VoIP fraud. The actor supplied it with a fraud knowledge base and asked for practical guidance on setting up a fraudulent Voice over IP operation. In a conversation, the actor asked about scams or methods used before, including call spoofing. With DeepSeek V4 underneath, the profile provided detailed criminal methodology, including advice on fraud-to-clean traffic ratios, carrier audit thresholds and avoiding detection.

A second profile, pandarouter, handled carrier research. The actor used it to examine a wholesale VoIP carrier for regulatory vulnerabilities and evidence of past fraud. The investigation found planning and setup activity, but no call logs, Session Initiation Protocol traces or active call routing. That absence leaves the operation’s status uncertain: the evidence points to preparation or an early build, rather than an active calling service.

The VoIP work shows how a model can act as a tutor for someone building expertise through conversation. That can shorten the path from curiosity to an actionable plan. Still, the available evidence does not establish that the actor completed the VoIP setup or used it to defraud anyone.

Wi-Fi cracking, credential scans and crypto research

The actor maintained an automated process for cracking Wi-Fi passwords. A Kali Linux device called Pi was placed within range of target networks to capture WPA authentication material, such as PMKID data or EAPOL handshakes. A server collected the captured material over an SSH tunnel, sent it to GPU instances on Vast.ai for cracking, then returned results to the actor through Discord webhooks.

This workflow depended on physical proximity to the targeted access points. That detail suggests the operator had local access to the locations, a requirement that cloud-based analysis alone cannot remove. The investigation does not say whether the recovered Wi-Fi credentials led to further access.

Investigators also found a large Go-based toolkit for scanning webpages and extracting exposed credentials, plus a companion Python script that gathered domain lists for scanning. In another line of reconnaissance, the actor used the DeBank API to scrape the top 500 ETH holders and major token holders. The resulting intelligence files listed 45 cryptocurrency “whales” with net-worth figures. Unit 42 assessed that the information was probably gathered for future targeting, but found no follow-on activity during its research.

The range of tools matters as much as any one script. The operator combined network access, credential discovery and financial reconnaissance, while keeping the potential targets and eventual use of some collected information unresolved.

A low-friction model stack

BUGSTER tested multiple language models, including GLM 5.2 from Zhipu AI and other models accessed through OpenRouter. The actor wrote scripts to test prompt overrides and flag refusals. Although GLM 5.2 is an open-weight model, the dedicated refusal checks show that the actor encountered safety barriers and tried to work around them.

Ultimately, the operator chose DeepSeek V4. Unit 42 links that choice to the model’s low cost and ease of access, as well as its susceptibility to a simple system-prompt override. The actor also configured Cursor through a proxy script called cipher-proxy.sh, which set environment variables to enable full agent mode and bypass permission and sandbox controls. The configuration and AI-generated scripts on the host indicate that Cursor was used for software development.

A single US-based server hosted several core services, while the Hermes profiles connected the operator to AI tools through Telegram. The precise service list and individual model configurations are not detailed in the source. Unit 42 says other, unrelated operators have adopted the Hermes and DeepSeek combination, including activity described in its earlier reporting on a Chinese-speaking actor. That overlap suggests this toolchain may be gaining traction among criminals seeking accessible AI support.

What the case shows—and what remains unclear

Unit 42 assesses BUGSTER as a financially motivated, English-speaking operator with no prior reporting. The actor’s writing was informal and often grammatically broken. Stored memory in the alice profile described the operator as a “gamer, OSINT, VoIP entrepreneur” and claimed the actor held a US LLC and EIN. Memory in pandarouter offered what investigators considered a more candid account, though the source provides no further attribution details.

The clearest demonstrated impact is the database intrusion and exposure of more than 5,200 student records. The investigation also documents attack-surface mapping and a candidate credential, while other activity—VoIP fraud and cryptocurrency targeting—remained at the planning or reconnaissance stage. That distinction is useful: AI can accelerate preparation and widen an operator’s reach, but tool configurations and plans alone do not prove that every intended attack succeeded.

For defenders, the case points to the value of watching for the results of these workflows: unusual endpoint enumeration, automated credential attempts, exposed secrets, unexpected database access, and suspicious Wi-Fi authentication capture. The next question is whether the Hermes and DeepSeek setup becomes a repeatable pattern across more operators, and whether future cases show planned activity progressing into completed fraud or access.

Unit 42’s related research on the Hermes and DeepSeek toolchain provides further context on that activity.