DeepSeek Empowers Hackers in Coordinated Cyberattacks Across 48 Countries, Print Vulnerability Emerges as Prime Example of AI Abuse

Overview

Cybersecurity company GreyNoise released its latest report, Agents Gone Wild, on September 9, revealing a large-scale cyberattack assisted by artificial intelligence. A hacker deployed hundreds of AI agents, combining the Codex framework with DeepSeek large language models, to launch intensive attacks against the global education sector. The attack exploited a recently disclosed vulnerability in Print software vendor PaperCut, affecting 395 organizations across 48 countries, and has become a textbook example of AI technology being abused for automated attacks.

Attack Scale and Speed

The automation level of this attack was extremely high, with the attacker successfully compromising 11 organizations in just 26 seconds. According to GreyNoise analysis, most of the affected victims were educational institutions, with one case at a US high school being particularly shocking: it took the attacker only 7 minutes from gaining initial access to taking over the domain administrator account.

Domain dump data generated by the attack came from 280 victims, demonstrating exceptionally high infiltration efficiency. Although the attacker configured an exclusion list covering 28 countries to avoid specific targets, the AI agents did not strictly adhere to these instructions during execution. Instead, they ignored the restrictions and continued to expand the attack scope. When Blackpoint recovered the attacker's state files, it was found that the operator was still modifying their deny list during the campaign, further confirming the potentially uncontrollable behavior of AI agents in complex tasks.

Root Cause Analysis: Lack of Non-Human Identity Governance

Industry experts widely point out that the PaperCut vulnerability is merely a surface symptom; the real underlying risk lies in the lack of governance over non-human identities within enterprises. For a long time, many companies have treated print or service accounts as static assets that are "set and forgotten," granting them persistent domain administrator privileges by default.

GreyNoise's report presents all CTOs with an urgent choice: continue to allow these privileged accounts to exist, or manage machine identities in the same way human employees are managed? Experts recommend adopting a short-lived token mechanism, eliminating persistent domain administrator privileges, and revoking access immediately after projects conclude. Only by addressing the governance gaps surrounding non-human identities can the spread of such AI-assisted attacks be fundamentally curbed.