Chinese Hacker Uses DeepSeek to Attack 460 Systems After Claude and OpenAI Blocked Offensive Use

My take: The Unit 42 (Palo Alto Networks) report published on July 30 confirms with concrete data that AI model safety controls have real operational value. A China-based actor tried using Claude and OpenAI models to automate cyberattacks, and both blocked him. Only when he turned to DeepSeek did he find a model willing to operate without those restrictions.

With a single Telegram command, the agent autonomously scanned, researched vulnerabilities, and attacked more than 460 systems. It successfully compromised three of them. The operation was exposed when the agent itself made a technical mistake that revealed its entire infrastructure.

For anyone making technology decisions at their company, this is a clear signal: choosing which AI provider you work with is no longer just a decision about functionality or price. Does your organization have a policy on which AI models can access your most sensitive systems?