OpenAI Disrupts China-Linked Cyber Operations Vixen and Keyhole Panda

OpenAI has identified and disabled accounts associated with China-linked threat actors known as Vixen and Keyhole Panda. These actors utilized OpenAI models to assist with technical reconnaissance, software development, and the automation of cyber operations, though OpenAI found no evidence that the models provided capabilities that were not already available through public resources.

Technical Behaviors and Operational Workflows

Threat actors engaged with OpenAI models in both English and Chinese, splitting their activities into two primary subsets of behavior:

Technical Research and Reconnaissance

One subset of activity focused on open-source research into specific entities and technical topics. This included using models to troubleshoot system configurations and modify scripts. Specific tools mentioned in these interactions included:

• reNgine: An automated reconnaissance framework for web applications.
• Selenium automation: Used to bypass login mechanisms and capture authorization tokens.

Infrastructure and Support Development

Another subset focused on support activities, including Linux system administration, infrastructure setup, and software development. Key activities included:

• System Administration: Configuring firewalls, nameservers, and building software packages for offline deployment.
• Software Development: Developing web applications, Android apps, and software written in C and Golang.
• Local LLM Deployments: Setting up infrastructure for local LLM deployments, such as DeepSeek, as well as configuring VPNs and Docker container deployments.

Malicious Use Cases and Content Generation

The threat actors generated content across several high-risk categories, ranging from network penetration to strategic research:

• Password Bruteforcing: Writing scripts to attempt multiple username and password combinations against FTP servers.
• Port Scanning: Modifying and improving scripts to scan servers for specific ports.
• AI-Driven Penetration Testing: Automating penetration testing by using LLMs to analyze Nmap scan output and iteratively generate new commands.
• Social Media Automation: Developing code to manage fleets of Android devices for automated social media operations.
• Strategic Research: Seeking publicly available information on US Special Operations Command, satellite communications, ground station terminal locations, government identity verification cards, and networking equipment.

Mapping to LLM ATT&CK Framework

OpenAI mapped the observed activities to the LLM ATT&CK framework to categorize the operational techniques used by the actors:

Technique	Activity
LLM Assisted Vulnerability Research	Researching vulnerabilities, generating AI-assisted penetration testing scripts via OpenAI API, and analyzing vulnerability reports to generate exploit payload ideas.
LLM Enhanced Scripting Techniques	Automating IP range conversion, network reconnaissance scripting, and automating command and control for Android social media manipulation.
LLM Guided Infrastructure Profiling	Profiling network infrastructure by extracting IPs and hostnames from pasted text.
LLM-Advised Strategic Planning	Analyzing telecom infrastructure and requesting details on government identity verification.
LLM-Optimized Payload Crafting	Generating code obfuscation and anti-reverse engineering techniques for malware development.
LLM Assisted Vulnerability Research	Researching vulnerabilities and generating AI-assisted penetration testing scripts.

Impact and Mitigation

OpenAI disabled all accounts associated with these activities and shared relevant indicators with industry partners. While the investigation provided visibility into the operational workflows of PRC-affiliated threat actors, OpenAI concluded that access to their models did not provide these actors with novel capabilities or directions that they could not have obtained from other publicly available resources.