BUGSTER Uses DeepSeek V4 and Hermes Agent for AI-Assisted Intrusions, Stealing 5,200+ Records from Indian University
Unit 42 reports that a threat actor using the handle BUGSTER set up DeepSeek V4 as the reasoning engine of an AI-assisted cybercrime operation against organizations in India, including universities. AI-generated reconnaissance and scripts were used to brute-force a university website. In a separate intrusion, the actor exfiltrated a MongoDB database with more than 5,200 student and faculty records containing PII. The activity shows how AI tooling can raise the capability of a relatively inexperienced operator across credential theft, data theft, wireless attacks, cryptocurrency research and fraud planning. Unit 42 links the case only thematically to its July 2026 report on another actor using Hermes Agent and DeepSeek; the campaigns are distinct.
Detalhes técnicos
- BUGSTER ran the Hermes Agent framework with DeepSeek V4, used Cursor as an additional AI coding tool, and controlled specialized Hermes profiles via Telegram ("alice" for VoIP fraud advice, "pandarouter" for carrier and call-routing research). The actor also tested other AI models and used prompt-level overrides to reduce safety restrictions.
- University attack: AI-generated reports mapped 12 subdomains (LMS, student portal, administration dashboard). WordPress REST API data exposed 13 user accounts. A generated password-guessing tool (117 usernames, 28 common passwords) yielded a valid login.
- Separate intrusion: an administrator account with a weak password gave access to a MongoDB database, which was exfiltrated (5,200+ records).
- Other tooling found: credential-scanning tools for exposed secrets (cloud and private keys), cryptocurrency-holder reconnaissance scripts, and a C2 listener supporting reverse shells.
- Wi-Fi cracking pipeline: a nearby Linux device captured authentication material and sent it to cloud GPU instances for cracking, suggesting physical proximity to targets (locations not identified).
- VoIP fraud: the AI chats covered business identity setup, carrier access, call routing and caller-ID evasion, including a "NEXT EXPLOIT OPPORTUNITIES" section. Unit 42 found no call logs, SIP traces or active routing, so this indicates planning or early setup, not confirmed fraud or successful STIR/SHAKEN evasion.
Remediação
The article gives no explicit remediation. The intrusions relied on weak and guessable credentials and exposed WordPress REST API user data, so strong unique passwords, MFA for admin accounts, limiting user enumeration and securing MongoDB administrator access are relevant.
Linha do tempo
-
BUGSTER Uses DeepSeek V4 and Hermes Agent for AI-Assisted Intrusions, Stealing 5,200+ Records from Indian University
ler a análise desta fonte ocultar
Unit 42 reports that a threat actor using the handle BUGSTER set up DeepSeek V4 as the reasoning engine of an AI-assisted cybercrime operation against organizations in India, including universities. AI-generated reconnaissance and scripts were used to brute-force a university website. In a separate intrusion, the actor exfiltrated a MongoDB database with more than 5,200 student and faculty records containing PII. The activity shows how AI tooling can raise the capability of a relatively inexperienced operator across credential theft, data theft, wireless attacks, cryptocurrency research and fraud planning. Unit 42 links the case only thematically to its July 2026 report on another actor using Hermes Agent and DeepSeek; the campaigns are distinct.
Technical Details
- BUGSTER ran the Hermes Agent framework with DeepSeek V4, used Cursor as an additional AI coding tool, and controlled specialized Hermes profiles via Telegram ("alice" for VoIP fraud advice, "pandarouter" for carrier and call-routing research). The actor also tested other AI models and used prompt-level overrides to reduce safety restrictions.
- University attack: AI-generated reports mapped 12 subdomains (LMS, student portal, administration dashboard). WordPress REST API data exposed 13 user accounts. A generated password-guessing tool (117 usernames, 28 common passwords) yielded a valid login.
- Separate intrusion: an administrator account with a weak password gave access to a MongoDB database, which was exfiltrated (5,200+ records).
- Other tooling found: credential-scanning tools for exposed secrets (cloud and private keys), cryptocurrency-holder reconnaissance scripts, and a C2 listener supporting reverse shells.
- Wi-Fi cracking pipeline: a nearby Linux device captured authentication material and sent it to cloud GPU instances for cracking, suggesting physical proximity to targets (locations not identified).
- VoIP fraud: the AI chats covered business identity setup, carrier access, call routing and caller-ID evasion, including a "NEXT EXPLOIT OPPORTUNITIES" section. Unit 42 found no call logs, SIP traces or active routing, so this indicates planning or early setup, not confirmed fraud or successful STIR/SHAKEN evasion.
Remediation
The article gives no explicit remediation. The intrusions relied on weak and guessable credentials and exposed WordPress REST API user data, so strong unique passwords, MFA for admin accounts, limiting user enumeration and securing MongoDB administrator access are relevant.