Security researchers discovered ZombieAgent, a technique exploiting ChatGPT's third-party app connections to create persistent backdoors and exfiltrate data. Attackers hide malicious prompts in emails, documents, or cloud storage that ChatGPT parses through its Connectors feature, enabling zero-click attacks with worm-like propagation capabilities. The vulnerability allowed manipulation of ChatGPT's Memory feature to create persistent data-leaking backdoors and even alter stored medical information. OpenAI patched these specific exploits in December 2024, but the fundamental prompt injection vulnerability remains without a complete fix.
Sort: