OpenAI Disrupts STORM-0817 Iran-Linked Malware Activity

OpenAI has disrupted a network of accounts operated by STORM-0817, an Iran-based threat actor. This marks the first time this specific actor has been publicly identified as using AI models to support its malicious operations.

Android Malware Development and C2 Infrastructure

STORM-0817 used OpenAI models to debug and implement Android malware and its associated command and control (C2) infrastructure. The malware, which was relatively rudimentary, possessed standard surveillanceware capabilities designed to retrieve the following data from compromised devices:

Contacts and call logs
Installed packages
Media and files (including PDFs and Excel documents) from external storage
Screenshots
Device IMEI and model
Browsing history
Latitude and longitude
Content downloaded to external storage, including files from secure messaging apps like WhatsApp and IMO

This code was implemented in two Android packages: com.example.myttt and com.mihanwebmaster.ashpazi.

To manage these compromised devices, the actor used ChatGPT to develop server-side code for a C2 server based on a WAMP (Windows, Apache, MySQL & PHP/Perl/Python) setup. During testing, this infrastructure utilized the domain stickhero[.]pro.

Reconnaissance and Scraping Operations

STORM-0817 employed AI models for reconnaissance and data collection against specific targets:

Instagram Scraping: The actor sought debugging help for Python code using the Selenium webdriver to scrape Instagram profiles. This tool was designed to retrieve details from all followers of a given username. Testing was observed on an Iranian journalist critical of the Iranian government.
Targeted Reconnaissance: The actor used models to translate the LinkedIn profiles of academics and individuals working at the National Center for Cyber Security in Pakistan into Persian. This activity aligned with reconnaissance prompts focusing on the National Cybercrime and Forensics Lab at the Air University in Pakistan, which supports the Pakistan military's Air Force, Navy, and Police.

Technical Mapping to LLM ATT&CK Framework

OpenAI categorized the actor's interactions with the models according to the LLM ATT&CK Framework:

Activity	LLM ATT&CK Framework Category
Debugging and implementing an Instagram scraper	LLM-enhanced scripting techniques
Translating LinkedIn profiles of Pakistani cybersecurity professionals	LLM-informed reconnaissance
Debugging and development support for Android malware and C2 infrastructure	LLM-aided development

Impact and Indicators of Compromise (IOCs)

OpenAI disabled all identified accounts associated with STORM-0817 and shared relevant Indicators of Compromise (IOCs) with industry partners.

Regarding the capability shift, OpenAI noted that the models provided only limited, incremental capabilities for these malicious tasks, as the goals were achievable using publicly available, non-AI powered tools.

Server-Side Indicators