OpenAI Disrupts SweetSpecter China-Linked Cyber Activity
OpenAI has identified and banned accounts associated with SweetSpecter, a suspected China-based adversary that attempted to leverage ChatGPT for offensive cyber operations while simultaneously targeting OpenAI employees and global governments with spear phishing attacks. This marks the first publicly identified instance of SweetSpecter targeting a U.S.-based AI company.
Spear Phishing Campaign Against OpenAI Employees
In May 2024, OpenAI received intelligence indicating that SweetSpecter was targeting corporate and personal email accounts of its employees. The adversary posed as a ChatGPT user seeking support to lure victims into opening a malicious attachment named ‘some problems.zip’.
Technical Execution of the Attack
The attack chain utilized a malicious LNK file to deliver a payload:
- Lure: The user is presented with a DOCX file containing simulated ChatGPT error and service messages.
- Payload: In the background, the SugarGh0st RAT (Remote Access Trojan) is decrypted and executed.
- Objective: The malware is designed to grant the adversary remote control over the compromised machine, enabling the execution of arbitrary commands, data exfiltration, and screen capturing.
OpenAI reports that existing security controls successfully blocked these emails before they reached corporate inboxes.
LLM-Based Tactics, Techniques, and Procedures (TTPs)
OpenAI disrupted a cluster of ChatGPT accounts using shared infrastructure to conduct vulnerability research and scripting tasks. These activities were mapped to the LLM-themed tactics, techniques, and procedures (TTPs) proposed for the MITRE ATT&CK® Framework.
Observed Adversary Activities
| Activity | LLM ATT&CK Framework Category |
|---|---|
| Inquiring about application vulnerabilities and specific CVE numbers | LLM-informed reconnaissance |
| Searching for vulnerable versions of Log4j (Log4Shell) | LLM-informed reconnaissance |
| Researching popular foreign content management systems | LLM-informed reconnaissance |
| Inquiring about the creation of internet-wide scanners | LLM-informed reconnaissance |
| Using sqlmap to upload web shells to target servers | LLM-assisted vulnerability research |
| Seeking exploits for a prominent car manufacturer's infrastructure | LLM-assisted vulnerability research |
| Programming communication services to send text messages | LLM-enhanced scripting techniques |
| Debugging extensions for cybersecurity tools | LLM-enhanced scripting techniques |
| Developing frameworks for programmatic SMS delivery | LLM-aided development |
| Researching social engineering themes for government employees | LLM-supported social engineering |
| Creating variations of job recruitment lure messages | LLM-supported social engineering |
Impact and Defensive Insights
OpenAI concluded that the use of LLMs did not provide the adversary with novel capabilities or directions that they could not have obtained from other public resources.
Use of AI in Defense
During the investigation, OpenAI security teams used ChatGPT to analyze, categorize, and summarize large datasets of adversary interactions. The company expects to use future iterations of the models to assist in reverse engineering and analyzing malicious attachments.
Importance of Collaboration
OpenAI emphasizes that collaboration with industry partners and the sharing of threat intelligence were critical in identifying the failed compromise attempts, highlighting the necessity of collective defense in the age of AI.