The Shared Clipboard Inside the Sandbox: Cross-Account Data Leakage in ChatGPT (research.checkpoint.com)

Check Point Research has uncovered a significant security vulnerability within ChatGPT that allows for cross-account data leakage. This exploit enables attackers to utilize a victim's ChatGPT session to perform unauthorized tasks, such as accessing and exfiltrating data from connected applications like Gmail, without the victim's knowledge. The vulnerability arises from a covert communication channel established between the isolated code-execution containers of different ChatGPT accounts, allowing attackers to embed instructions that the model processes alongside the victim's requests. Even while delivering regular responses, ChatGPT can execute hidden tasks driven by the attacker’s instructions, effectively making it a coerced insider. This discovery highlights critical implications for the AI/ML community regarding the security architecture of large language models (LLMs) that can interact with external data and tools. As AI systems gain more capabilities, their operational security must evolve to protect user data within this increasingly complex landscape. The incident underscores the necessity for stronger isolation mechanisms and thorough audits of internal communication layers within AI platforms. Although OpenAI confirmed the compromised internal service has been decommissioned, this vulnerability serves as a cautionary tale about the risks posed by shared environments in cloud-based AI applications, which necessitate ongoing vigilance and enhanced security protocols.