Lone criminal used AI agents to attack Indian universities, Unit 42 says

A financially motivated criminal using the handle BUGSTER relied on AI agents with their safety controls turned off to plan and carry out attacks, mostly in India, Unit 42 says. With AI-written tools, the attacker ran a password attack on one Indian university’s WordPress site and copied a database of more than 5,200 records from a second. The records include students’ personal data. Unit 42 says the case shows how AI lowers the skill needed for cybercrime.

Unit 42 says the attacker ran the Hermes Agent framework with DeepSeek V4 as its model, controlled through Telegram. The attacker also used the Cursor code editor with its permission checks and sandbox turned off. Separate AI personas gave advice on internet phone fraud and carrier research. A pipeline cracked Wi-Fi passwords captured near targets, so the attacker was physically close to some of them.

At the first university, the attacker’s AI-written reports listed 12 subdomains. The attacker then pulled 13 user IDs with real names from post author data and author sitemaps, getting past a Cloudflare firewall rule that blocked direct requests. A generated tool tried 117 usernames with 28 common passwords and flagged one login as valid. At the second university, an administrator account with a weak password exposed a MongoDB database.

Unit 42 links the attacker to one US server, 162.243.3[.]92, that hosted the AI agent, a command listener and other tools. It says other, unrelated attackers have taken up the same pairing of Hermes Agent and DeepSeek, which it calls cheap, easy to reach and easy to jailbreak.