OpenAI has banned accounts identified as belonging to CyberAv3ngers, a threat actor publicly reported to be affiliated with Iran’s IRGC. These accounts utilized OpenAI's models for various research activities related to industrial control systems (ICS), default credentials, and potential targets. This case study was initially published in OpenAI’s October 2024 report.
Key Points
- OpenAI banned accounts appearing to belong to CyberAv3ngers based on assessment from a credible source.
- The threat actor used OpenAI models to research vulnerabilities, debug code, and seek scripting advice.
- CyberAv3ngers is known for disruptive attacks against ICS and programmable logic controllers (PLCs) in water, manufacturing, and energy systems.
- Targeted infrastructure is typically associated with Israel, the United States, or Ireland.
- Observed activities included reconnaissance, asking for default username/password combinations for PLCs, and creating/refining bash and Python scripts.
- Requests sometimes indicated an interest in or targeting of Jordan and Central Europe.
- OpenAI believes these interactions provided only limited, incremental capabilities, not novel ones, achievable with publicly available tools.
Context
According to OpenAI, the CyberAv3ngers group has been linked to incidents such as the compromise of PLCs at the Municipal Water Authority of Aliquippa, Pennsylvania, in November 2023, and a two-day disruption of water services in County Mayo, Ireland, in December 2023. These campaigns often exploit default or weak passwords and documented PLC vulnerabilities, using open-source tools for scanning and exploitation.
Why It Matters
This action highlights how AI model providers are monitoring and responding to malicious use by state-sponsored actors. It demonstrates that deployers of AI models must consider the potential for misuse in reconnaissance and vulnerability research, even if the capabilities provided are incremental.
What To Do
- Note that OpenAI's October 2024 report contains further details on this case study.
- Observe the types of queries identified, such as requests for default credentials or vulnerability information, as indicators of potential misuse.
- Review the examples of specific queries made by the threat actor to understand the scope of their AI-assisted research.
- Watch for further reports from AI model providers regarding their actions against malicious state-sponsored activities.