AI-Assisted Cybercrime: One Operator’s Use of Hermes, DeepSeek and Cursor

One operator used AI assistants to work across several kinds of cybercrime, from university network intrusions to planned VoIP fraud. The investigation by Palo Alto Networks’ Unit 42 shows how a small set of tools can help a single person research targets, generate code and seek specialized advice. In one case, the operator stole a database containing more than 5,200 records with students’ personal information.

The operator, who used the handle BUGSTER, built a setup around Hermes Agent, with DeepSeek V4 as its reasoning model and Cursor as a coding assistant. Unit 42 found that the operator disabled safety controls in both AI tools. The setup also included distinct bot profiles for different tasks, connected through Telegram.

University targets and stolen records

One effort focused on an Indian university’s online education platform. Unit 42 found AI-generated reconnaissance reports that organized the target’s exposed routes and assigned severity ratings. The operator then attempted to brute-force access to the site. The investigation describes this as a four-stage sequence, from reconnaissance through credential attempts, and found a recorded candidate credential.

A separate intrusion affected another Indian university. The operator extracted a MongoDB database containing more than 5,200 records, including personally identifiable information about students. The source does not give a full account of the records’ contents or explain how the intrusion was achieved.

Unit 42 also found evidence that the operator enumerated a site running a popular e-commerce platform and collected information from it. The resulting report appears to have been generated with a large language model. It included a section titled “NEXT EXPLOIT OPPORTUNITIES,” which laid out possible follow-up activity. That document offers a useful glimpse of AI’s role here: the system helped turn collected observations into an organized plan.

VoIP fraud research remained at an early stage

The operator created a Hermes profile called alice to build knowledge of VoIP fraud. Unit 42 found conversations in which the bot supplied detailed criminal guidance, alongside a fraud knowledge base the operator had added to the profile. The operator also used a second profile, pandarouter, to research wholesale VoIP carriers.

The discussions covered fraud traffic patterns, carrier scrutiny and ways to avoid detection. Unit 42 says the guidance included suggested ratios of fraudulent to legitimate traffic, audit thresholds and operational precautions. The source does not publish the specific values, and the available evidence does not show that the advice was put into practice.

That distinction matters. Investigators found no call logs, Session Initiation Protocol traces or active call routing to show that a working VoIP operation was running. Their assessment is that the effort was still in the planning or early setup phase. The AI bots could supply a form of on-demand domain expertise, but the investigation found evidence of planning rather than completed VoIP fraud.

Automated collection and cracking tools

The operator also maintained an automated process for cracking Wi-Fi passwords. A Linux device, labeled Pi, captured authentication material from nearby wireless networks. The captured data travelled through an SSH tunnel to a server, which sent it to cloud GPU instances on Vast.ai for cracking. Results were returned through Discord webhooks.

This method depends on being close enough to the target network to capture its authentication exchange. Unit 42 therefore assesses that the operator had local access to the locations being targeted. The investigation does not identify those locations or describe how access was obtained.

Other tools gathered credentials exposed on webpages. Unit 42 found a large toolkit written in Go, plus a Python script that collected domain lists for scanning. The source does not provide a count of the credentials collected or identify any resulting account takeovers.

For cryptocurrency research, the operator queried the DeBank API for leading ETH holders and major token holders. The resulting files described 45 large holders and included estimated net worth figures. Unit 42 believes the material was gathered for possible future targeting, but found no follow-on activity during its research.

One server, several AI profiles

The operator hosted several core services on a single US-based server, identified in the investigation as 162.243.3[.]92. Hermes Agent was the main interface, controlled through Telegram direct messages. Three profiles ran as separate bot gateways, with alice and pandarouter assigned distinct roles. Four more profiles—dev, check, vuln and deploy—appeared in a shared Telegram group. Unit 42 found no conversations for those four, though their names suggest intended uses in coding, checking, vulnerability work and deployment.

Cursor was configured through a proxy script named cipher-proxy.sh. The script set environment variables that enabled agent mode and bypassed permission and sandbox controls. Unit 42 says the configuration and AI-generated scripts on the server indicate that Cursor was used to help develop code.

Logs and test scripts show that BUGSTER tried several models, including GLM 5.2 from Zhipu AI and models reached through OpenRouter, before settling on DeepSeek V4. The operator wrote scripts to test prompt overrides and record when GLM 5.2 refused requests. Unit 42 describes GLM 5.2 as an open-weight model that had been reported as weakly guarded, but the refusal checks suggest the operator still encountered resistance.

The chosen setup used a system prompt to try to override model safeguards. Unit 42 links this choice to its earlier reporting on other activity using Hermes with DeepSeek, and says multiple unrelated operators have adopted the same pairing. Its assessment is that low cost, ease of access and limited resistance to prompt overrides help explain the combination’s appeal.

What this case shows

Unit 42 characterizes BUGSTER as an English-speaking, financially motivated operator working across several criminal areas. Notes stored in the bots described the operator as a “gamer, OSINT, VoIP entrepreneur”; one profile also claimed the operator held a US LLC and EIN. The investigation says there was no prior reporting on this individual.

The practical change is less about AI autonomously carrying out an intrusion than about one operator using assistants to span tasks that usually call for different kinds of experience. Here, AI helped organize reconnaissance, produce scripts and provide specialized advice, while conventional infrastructure handled data collection and processing. The outcome included a confirmed university intrusion and stolen student records; other activity, including VoIP fraud and cryptocurrency targeting, remained at the planning or research stage in the evidence described.

Unit 42’s findings point to a toolchain that can lower the effort needed to move between technical and operational tasks. For defenders, the case makes it useful to watch for the combination of AI-generated reconnaissance, automated credential collection and unusual data handling. The next question is whether operators using similar setups can turn exploratory plans into repeatable activity—and whether model providers can make simple prompt overrides less effective.