Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Unit 42 header image of hands typing on a laptop keyboard beneath a glowing AI chip graphic and lines of light.
Photo illustration: Unit 42, Palo Alto Networks, header image for its BUGSTER research (stock image, Adobe Stock), Oct 21, 2026 (fair use).

Threat Intel

Unit 42 Says a Low-Skill Actor Called BUGSTER Ran Jailbroken Hermes Agent and DeepSeek V4 Bots to Attack Indian Universities, Exfiltrate More Than 5,200 Records and Plan a VoIP Fraud Operation

Palo Alto Networks' threat unit says the financially motivated operator, with no prior public reporting, ran its AI assistants from a single U.S. server, steered them over Telegram and turned off the safety controls in both Hermes and the Cursor coding tool. Unit 42 also says multiple unrelated actors are now adopting the same Hermes and DeepSeek pairing. The indicators and what the playbook means for defenders.

Priya Shah, Threat IntelSingapore6 min read

SINGAPORE - A cybercriminal operating under the handle BUGSTER leaned on jailbroken AI assistants to map targets, write attack tools and learn a fraud business, and still came away with personal data on thousands of university students and staff, according to research Palo Alto Networks' Unit 42 published Wednesday.

Unit 42 said it identified and investigated the actor's operational environment, which gave it a direct view of how the tools were set up and used. The actor configured the Hermes Agent framework with DeepSeek V4 as its reasoning engine, controlled through Telegram direct messages, and the Cursor integrated development environment as a secondary tool, "both with safety controls disabled."

The attacker "used basic tooling and relied on AI-generated scripts and reconnaissance reports, suggesting they didn't have the skills these activities would have required before the advent of readily available AI tools," Unit 42 wrote. One exchange it quoted shows the register: "Can you find about like what scam or thing they done in past like spoofing calls or somthing [sic]."

Unit 42 flowchart of the actor's main server linked to a proxy, target organizations, the DeepSeek API, Hermes Agent profiles, Telegram, Cursor IDE and a Wi-Fi cracking pipeline.
Chart: Unit 42, Palo Alto Networks, the actor's infrastructure and AI tool relationships, Oct 21, 2026 (fair use).

Two universities

Against an Indian university's online education platform, the actor's AI-generated reconnaissance found 12 subdomains, including the main WordPress site, a learning management system, a student portal and an administration dashboard. Enumeration through the WordPress REST API pulled 13 active user IDs with real names from post author fields and author sitemaps, getting around the Cloudflare web application firewall that blocked direct access to the user endpoint. A generated brute-force tool then paired 117 usernames with 28 common passwords, and it flagged one WordPress login as valid.

Unit 42 screenshot of the actor's WordPress REST API enumeration report, with an executive summary rating the risk medium and a high-severity finding that 13 active user IDs and real names are exposed, names obscured.
Screenshot: Unit 42, Palo Alto Networks, the actor's AI-formatted enumeration report, Oct 21, 2026 (fair use).

At a second Indian public university, the actor exfiltrated a MongoDB database. The tooling's manifest recorded an administrator account with a weak password, student records at diploma, undergraduate and postgraduate level and faculty records containing personal data, plus details of the university's cloud load balancing, edge routing, database services and a third-party cloud storage service account. Unit 42 put the haul at more than 5,200 records.

The actor also enumerated an Indian store on a popular e-commerce platform, extracting a storefront API token, three admin email addresses, two phone numbers and payment merchant IDs. The resulting report, which Unit 42 suspects was written by a large language model, carried a section headed NEXT EXPLOIT OPPORTUNITIES.

An AI fraud adviser

Three Hermes profiles ran as separate Telegram bots: make-good, a general-purpose criminal assistant with safety overrides; alice, a VoIP fraud adviser; and pandarouter, for route management and carrier research. Four more, named dev, check, vuln and deploy, sat in a shared Telegram group. With DeepSeek V4 underneath, alice laid out a plan that follows an established model: form a U.S. limited liability company in Wyoming or Delaware as a non-resident, obtain an employer identification number without a Social Security number by fax to the IRS, complete FCC registration, buy SIP trunk capacity from legitimate carriers and route roughly 70 percent legitimate traffic alongside 30 percent fraudulent calls to stay below carrier audit thresholds, while getting around the STIR/SHAKEN caller ID framework. Unit 42 found no call logs, SIP traces or active call routing, suggesting the operation was still in planning or early setup.

Unit 42 excerpt of a conversation in which the actor asks about running 70 percent clean traffic or 30 percent scam traffic and the alice bot explains a 70/30 split and warns about carrier audits.
Screenshot: Unit 42, Palo Alto Networks, an excerpt from the actor's alice conversation, Oct 21, 2026 (fair use).

One U.S.-based server, 162.243.3.92, hosted the Hermes framework, a Python command-and-control listener for Windows reverse shells and a Cursor proxy. A script named cipher-proxy.sh set environment variables to force Cursor's agent mode and bypass its permission and sandbox controls. The actor tried GLM 5.2 and other models reached through OpenRouter, wrote a script to send a safety override to GLM 5.2 and flag refusals, and settled on DeepSeek V4.

The work went beyond web intrusions. A Kali Linux device the actor labeled Pi, placed within range of target networks, captured WPA handshake or PMKID material, which the server collected over an SSH tunnel and dispatched to Vast.ai cloud GPU instances for cracking, with results reported through Discord webhooks. That requires physical proximity, which Unit 42 said indicates local access to the targeted locations. A Go toolkit scraped web pages for exposed .env values, S3 signed URLs, Google Cloud service account keys, CDN API keys and private keys in log files, and the actor scraped the top 500 ETH holders through the DeBank API, producing files on 45 cryptocurrency whales.

A toolchain, not one actor

"Multiple, unrelated actors are independently adopting the same Hermes + DeepSeek combination," Unit 42 wrote, pointing to its earlier report on a Chinese-speaking actor that used the same pairing. It assessed that a key driver is "low safety friction at low cost," because DeepSeek models are inexpensive, easy to access and "readily jailbroken with a simple system-prompt override."

Unit 42 assessed BUGSTER as a financially motivated, multi-domain criminal operator with no prior reporting, English-speaking with informal, frequently broken grammar.

What defenders can do

  • Sweep logs for 162.243.3.92, the actor's server, and 104.238.38.202, a commercial proxy.
  • Give fraud and abuse teams the attribution indicators: email bugstergap@gmail.com, display name B U G S T E R, Cursor IDE user ID 158034011 and Telegram user ID 7593151260.
  • WordPress operators should check whether post author fields and author sitemaps expose real user names even where the WAF blocks the users endpoint, as they did here, and alert on login bursts that walk a list of real user names.
  • Keep MongoDB and other databases off the internet and audit administrator accounts for weak passwords, the opening Unit 42 recorded at the second university.
  • Scan your own public web content for exposed .env files, signed storage URLs and cloud keys, the material the actor's Go toolkit was built to harvest.

Sources:


Priya Shah covers threat intelligence, intrusion analysis, and adversary tradecraft for SOCtember from Singapore.

Related stories

Threat Intel desk