AI-Assisted Cybercrime: One Operator Combined University Intrusions, Wi-Fi Cracking and VoIP Planning
- Unit 42
- 385 reads
A threat actor used Hermes Agent and DeepSeek V4 to coordinate intrusions and prepare other schemes, showing how AI can widen one operator’s reach.

One operator used AI assistants to move between university network intrusions, Wi-Fi password cracking, cryptocurrency research and plans for VoIP fraud. The case shows the practical value these tools can offer a criminal: help producing scripts and reconnaissance reports, plus conversational guidance in areas where the operator lacked experience.
Investigators at Unit 42 attributed the activity to an operator using the handle BUGSTER. Their analysis found a Hermes Agent setup connected to DeepSeek V4, with Cursor as a secondary coding tool. The operator disabled safety controls and accessed the assistants through Telegram. The result was a collection of separate AI profiles and automated services that supported different tasks, all run from infrastructure the operator controlled.
Two university targets, different outcomes
At one Indian university, BUGSTER used AI-assisted reconnaissance to map an online education platform and then ran a brute-force attempt. The reconnaissance reports were structured, with summaries, severity ratings and categories of endpoints. Investigators judged that style consistent with machine-generated output. Their account describes a four-stage progression, from mapping the site to testing credentials, and says the attempt recorded a candidate credential.
A separate intrusion against another Indian university had a more tangible result. BUGSTER extracted a MongoDB database containing more than 5,200 records, including students’ personally identifiable information. The extraction tools created a manifest alongside the dumped collections, documenting the captured data. The source material does not specify the individual fields in those records, so the scale and sensitivity of the information are clearer than its exact contents.
The actor also enumerated a site built on a popular e-commerce platform. The available findings describe that activity as data extraction, but don’t detail what information was obtained or whether it led to further access. In a separate suspected AI-generated report, a section titled “NEXT EXPLOIT OPPORTUNITIES” laid out potential follow-up attacks. That points to AI playing a role in organizing the operator’s next steps, as well as helping with technical execution.
AI as a specialist adviser
BUGSTER’s use of AI extended beyond writing code. The operator created specialized Hermes profiles, including one called alice, to build knowledge of VoIP fraud. That profile had a fraud knowledge base and gave detailed guidance through conversation. The reported topics included traffic ratios, carrier audit thresholds and ways to avoid detection.
The approach resembles a common VoIP fraud model, but the findings describe planning and early setup rather than an established calling operation. Investigators found no call logs, Session Initiation Protocol traces or active call routing. Another profile, pandarouter, was used for carrier research, including inquiries into a wholesale carrier’s vulnerabilities and history of fraud. The distinction matters: the conversations indicate intent and preparation, while the absence of call evidence limits what can be said about execution.
This is a different kind of assistance from code completion. An AI profile can answer follow-up questions, retain a focused body of reference material and help an operator reason through a sequence of tasks. In this case, investigators say that DeepSeek V4 supplied actionable criminal guidance. That may let a person with limited prior expertise attempt work that would otherwise require more research or help from a specialist.
Automation tied to physical access
The operator also built a Wi-Fi password-cracking pipeline. A Kali Linux device called Pi captured authentication material, such as PMKID data or EAPOL handshakes, from nearby access points. The server collected those captures over an SSH tunnel and sent them to GPU instances on Vast.ai for cracking. Results were returned through Discord webhooks.
Cloud compute handled the cracking work, but it could not remove the need to be near the networks during capture. Investigators therefore assessed that BUGSTER had local access to the targeted locations. That operational constraint is a useful reminder that AI and remote infrastructure can support an attack without making every part of it remote.
Elsewhere, a Go-based toolkit scanned web pages for exposed credentials, while a companion Python script gathered domain lists for scanning. The operator also scraped data on the largest ETH holders and major token holders through the DeBank API. The resulting files identified 45 cryptocurrency “whales” and included net-worth figures. Investigators considered the data likely intended for future targeting, but found no follow-on activity during their research.
A low-friction AI toolchain
Hermes Agent was the main interface, controlled through Telegram direct messages. Three profiles ran as separate Telegram bot gateways, each providing a specialized mode. Four more profiles—dev, check, vuln and deploy—were configured in a shared Telegram group. Investigators found no chat records for those profiles; their names suggest intended roles in development, validation, vulnerability testing and deployment.
Cursor was configured through a proxy script called cipher-proxy.sh. The setup enabled agent mode and bypassed permission and sandbox controls. Alongside AI-generated scripts found on the host, that configuration suggests Cursor supported code development.
BUGSTER tested several models before settling on DeepSeek V4. Logs and scripts also showed experiments with GLM 5.2 from Zhipu AI and models available through OpenRouter. A dedicated script sent GLM 5.2 a prompt intended to override its safety controls and recorded refusals. That detail shows the operator had to test which systems would comply; weaker safeguards were a sought-after feature, not an automatic property of every model.
Unit 42 says other, unrelated actors have adopted the Hermes and DeepSeek combination. Its earlier reporting described a Chinese-speaking operator using the same pairing. That suggests the setup may be becoming a repeatable choice for criminals seeking inexpensive, accessible models with limited resistance to prompt-level overrides. The findings don’t establish how common it is, but they point to a toolchain that other operators can assemble without building their own AI platform.
What the case establishes
Unit 42 assesses BUGSTER as financially motivated and active across several types of crime. The operator’s stored profile described them as a “gamer, OSINT, VoIP entrepreneur” and claimed a US LLC and EIN. Investigators also characterized the operator’s written English as informal and frequently ungrammatical. Those details provide context, but the technical evidence is more consequential: one person used AI to support reconnaissance, credential attacks, data theft, Wi-Fi cracking and planning for VoIP fraud.
The operator ran multiple core services from a single US-based server, listed by investigators as 162.243.3[.]92. It also hosted a Python command-and-control listener for Windows reverse shell connections. The report identifies supporting payloads but does not describe them in the supplied findings, so their capabilities cannot be assessed here.
The confirmed university database theft is the clearest measure of impact: more than 5,200 records were taken. Other activity varied from attempted access to research that investigators believe was intended to support later targeting. That mix makes it important to distinguish completed intrusions from preparation. The VoIP operation, in particular, had substantial planning evidence but no observed active routing.
For security teams, the case suggests watching for connected behaviors rather than treating each tool as an isolated clue: structured reconnaissance, repeated credential attempts, exposed-secret scanning, unusual remote compute use and bot-driven operational guidance. The wider question is whether AI profiles make it easier for one operator to sustain activity across several technical domains. DeepSeek V4 and Hermes Agent offer a concrete example; future observations will show whether this pairing persists or gives way to other low-cost combinations.