AI-Assisted Intrusion: DeepSeek-Powered Attacker Steals Over 5,200 University Records

A relatively unsophisticated threat actor using the handle BUGSTER leveraged AI tools to undertake a string of intrusions against Indian organizations, including at least two universities, and exfiltrated a MongoDB database with more than 5,200 records, according to researchers at Palo Alto Networks’ Unit 42. The case illustrates how readily available AI assistants, when combined with weak account hygiene and ordinary scripts, can amplify the capabilities of an attacker who lacks deep technical expertise.

What happened

Unit 42’s investigation links the activity to an operator who combined an AI assistant framework called Hermes Agent with DeepSeek V4 as the reasoning model and Cursor as a coding aid. In one confirmed breach, the attacker obtained an administrator account protected by a weak password and used it to access a public university’s MongoDB instance, stealing student and faculty records across diploma, undergraduate and postgraduate programs. Unit 42 did not name the affected universities.

The attacker’s methodology

The intrusions combined basic reconnaissance, automated credential scanning, and AI-generated guidance and scripts. In one instance the actor used AI-assisted research to map a university’s online education ecosystem — a WordPress site, a learning platform, a student portal and an administration dashboard — and then attempted password guessing until a valid login was found. Separately, another intrusion implicated a weak administrator password that directly led to database access and data theft. Unit 42 found evidence the attacker turned off safety controls in the AI tooling and relied on the model to produce reports and code.

Tools, preparation, and scope of activity

Beyond the two university incidents, investigators found artifacts showing the actor used a credential-scanning toolkit aimed at exposed secrets (cloud keys, private keys in logs/files), automated Wi‑Fi password cracking routines that required physical proximity, and profiles configured to research telecom fraud and large cryptocurrency holders. While the actor drafted plans and collected reconnaissance for schemes such as VoIP fraud, Unit 42 did not find definitive evidence those plans were executed — the report documents preparation and intent as well as confirmed theft in at least one case.

How AI lowered the bar

This case demonstrates that AI can serve as a force multiplier: it produced structured research, generated scripts, and helped organize multi-step attacks. But the report is clear that the model did not act alone; rather, the attacker used AI to accelerate tasks that previously required more manual skill. The combination of ordinary tools, exposed services, and weak account protections created the conditions for successful exploitation.

Recommendations for organizations

• Review and harden administrator accounts: remove weak or reused passwords and require multi-factor authentication wherever possible.
• Limit database access: restrict administrative access to approved systems and users, and avoid exposing management ports and databases to the public internet.
• Protect secrets: keep API keys, cloud credentials, and private keys out of public files and logs; rotate or revoke any credential that may have been exposed.
• Monitor for anomalous activity: watch for unusual logins, rapid credential-scanning behavior, or bulk access patterns that could indicate automated exfiltration.
• Manage AI agent permissions: if AI assistants are used in legitimate workflows, limit their permissions, keep safety controls enabled, and regularly audit agent actions.

Indicators of compromise (IoCs):-

Where this fits in the broader threat landscape

Reports like this sit alongside other documented DeepSeek-powered and AI-assisted cyber activity that raise concerns about model misuse. The key takeaway for defenders is not that AI is inherently malicious, but that easy access to decision-making assistants makes it far more feasible for lower-skilled attackers to orchestrate complex, multi-stage intrusions when basic security controls are missing.

Licensed under CC BY-ND 4.0 International