Low-skilled criminal used AI as an adviser for hacking attacks

Security firm Unit 42 (Palo Alto Networks) has described a criminal who calls himself BUGSTER and who used AI to plan and carry out several forms of cybercrime. According to Unit 42, he relied on AI-generated scripts and target mapping, and would not have had the necessary skills on his own before AI tools existed.

BUGSTER describes himself as a "gamer, OSINT, VoIP entrepreneur."

How he worked

He used Hermes Agent, an AI agent that can perform tasks on its own, with the language model DeepSeek V4 as its "brain." Everything was controlled through three Telegram bot profiles. As a secondary tool, he used the coding tool Cursor IDE.

Both setups had their safety controls bypassed with a so-called system prompt override, a hidden base instruction that makes the AI model ignore its own restrictions.

Two universities in India

At one university, AI-generated reports mapped 12 subdomains, including a learning system, a student portal and an administration dashboard. On the university's WordPress site, he extracted 13 user IDs and used an AI-generated tool to guess passwords with 117 usernames and 28 common passwords. This produced a valid login.

According to Cybersecuritynews.com, which relays Unit 42's findings, he also extracted a database with more than 5,200 records containing personal information about students and faculty. Access was gained through an administrator account with a weak password. This is not directly seen in Unit 42's own text.

The breaches exploited basic weaknesses such as exposed services, weak passwords and insufficient access control, not software flaws or zero-days.

Fraud and other plans

He also mapped an e-commerce platform and extracted API tokens and administrator emails. In addition, he received AI guidance on phone fraud over the internet (VoIP), including on setting up a company and bypassing the STIR/SHAKEN call verification. The AI advised him to mix roughly 70% legitimate traffic with 30% fraudulent traffic. According to Unit 42, the fraud was at the planning or early setup stage, and there are no confirmed active calls.

He also worked on cracking Wi-Fi passwords, mapping the 500 largest ETH holders and developing tools for controlling other people's computers.

Unit 42's assessment

Unit 42 assesses that AI functioned as an "on-demand criminal advisor," allowing him to gain operational expertise through conversation. The firm has seen several unrelated actors use the Hermes and DeepSeek combination and believes it is becoming a favorite among criminals. The reasoning is that DeepSeek models are cheap, easy to access and easy to jailbreak.

The assessments are the security vendor's own, and the analysis is based on the actor's own logs and test scripts, so the picture may be incomplete. Unit 42 could not see any follow-up activity against the crypto targets.