OpenAI Disrupts Deceptive IT Employment Schemes

OpenAI has banned ChatGPT accounts linked to multiple deceptive employment campaigns designed to fraudulently obtain remote IT and software engineering roles globally. These threat actors utilized AI to automate the creation of credible personas and resumes, and to research methods for bypassing corporate security and identity verification.

Threat Actor Profile and Attribution

OpenAI identified behaviors consistent with activity publicly attributed to North Korean (DPRK) linked IT worker schemes. The operation appears to have involved a tiered structure consisting of core operators and contractors. Some contractors may have been based in North America and were recruited to receive company laptops on behalf of the core group to facilitate remote access.

Operational Behavior and AI Integration

The threat actors integrated AI into every stage of the employment process to increase operational efficiency. This process was split between two types of operators:

Core Operators

Core operators focused on the infrastructure and automation of the fraud scheme. Their activities included:

• Automated Resume Generation: Using specific job descriptions, skill templates, and persona profiles to automate the creation of tailored resumes.
• Recruitment: Generating content to recruit contractors globally to act as proxies for job applications and hardware operation.
• Infrastructure Research: Using ChatGPT to research remote-work setups and the recruitment of US-based individuals to take delivery of corporate laptops for remote access.
• Security Evasion: Researching tools such as Tailscale peer-to-peer VPN, OBS Studio, vdo.ninja live-feed injection, and HDMI capture loops to circumvent corporate security and bypass identity verification during live video meetings.

Contractor Operators

Contractor operators used ChatGPT for tactical execution, including:

• Application Tasks: Completing job application requirements and coding assignments.
• Communication: Generating messages to core operators regarding payments and the personas used for the applications.

Technical Use Cases of LLMs

OpenAI categorized the malicious use of their models into four primary technical domains:

• LLM Supported Social Engineering: Automating detailed resumes aligned with industry norms and answering employment-related questions or coding assignments based on uploaded resumes.
• LLM-Enhanced Anomaly Detection Evasion: Seeking guidance on configuring corporate laptops to mask geolocation and evade endpoint security.
• LLM Aided Development: Coding tools to automate mouse movement or keep computers awake remotely to maintain a persistent presence on corporate networks.

Impact and Detection

While OpenAI cannot independently assess the success rate of these operations, the reliance on AI increased the threat actors' exposure. The detailed prompts and automation loops used by the actors provided OpenAI with insights into the full workflow of the campaigns, which have since been shared with industry peers and authorities to improve collective detection and response capabilities.