OpenAI Disrupts Cyber Operation Targeting Semiconductor and Academic Sectors

OpenAI Disrupts PRC-Linked Cyber Operation

OpenAI has banned a cluster of ChatGPT accounts used by threat actors to support cyber operations aligned with PRC intelligence requirements. The actors used the models primarily to increase the efficiency of existing workflows, specifically for crafting localized phishing content and debugging routine malware tooling, rather than to develop novel offensive capabilities.

Threat Actor Identification and Targeting

OpenAI identified the account cluster as overlapping with threat groups tracked by industry partners as UNK_DROPPITCH (Proofpoint) and UTA0388 (Volexity). In at least one instance, an email address used to register a ChatGPT account was also used to send phishing messages.

The operations displayed hallmarks of PRC intelligence requirements, characterized by:

  • Language Use: Extensive use of the Chinese language.
  • Targeting: Focus on Taiwan’s semiconductor sector, U.S. academia and think tanks, and organizations associated with ethnic and political groups critical of the CCP (referred to as the ‘five poisons’).

Behavioral Analysis of LLM Usage

Threat actors used ChatGPT to accelerate routine tasks and improve the quality of their social engineering and technical tooling. Their development capabilities were described as technically competent but unsophisticated, evidenced by the use of a simple static key despite discussing the nuances of AES encryption for C2 traffic.

Phishing and Social Engineering

The actors followed a formulaic playbook to create highly targeted phishing content in English, Japanese, and both simplified and traditional Chinese. They used the model to:

  • Generate concise, formally polite emails from academic, industry, or conference personas.
  • Adjust tone and swap terms for regional usage to increase authenticity.
  • Limitation: Despite these efforts, the actors failed to correct implausible example contact details in signature blocks.

Tooling and Malware Development

The actors requested code snippets and checklists to accelerate the development of low-to-mid maturity malware. Their activities included:

  • C2 Infrastructure: Drafting and repairing Go client-server code featuring AES-GCM encryption, session rekeying, and system-info beacons. They sought to move from plain WebSockets to secure WebSockets (wss://) and HTTP to HTTPS.
  • Process Control: Using Go and PowerShell to enumerate processes, gather environment details, and terminate specific Edge/WebView2 processes.
  • Reconnaissance: Using PowerShell for antivirus discovery and integrating commodity scanners like nuclei and fscan on commercial Linux infrastructure.
  • OPSEC: Requesting simple obfuscation tweaks, such as renaming functions, tweaking headers, and hiding strings to evade detection.

OpenAI noted that implementation details in the actors’ Go-based development overlapped with malware tracked as GOVERSHELL (Volexity) or HealthKick (Proofpoint).

Mapping to LLM ATT&CK Categories

The actors' activities map to several LLM ATT&CK categories, indicating a structured approach to utilizing AI for cyber operations:

  • LLM-Optimized Payload Crafting
  • LLM-Enhanced Anomaly-Detection Evasion
  • LLM-Assisted Post-Compromise Activity
  • LLM-Assisted Reconnaissance and Discovery
  • LLM-Assisted Social Engineering

Impact and Conclusion

OpenAI disabled all associated accounts and shared indicators with industry partners. The investigation concluded that the models did not provide the actors with novel offensive capabilities, as all activities relied on well-documented public techniques. The primary advantage gained by the actors was incremental efficiency in linguistic fluency, localization, and the speed of producing "glue code" and adjusting tools when failures occurred.

Sources