OpenAI has taken steps to address a data exfiltration vulnerability in ChatGPT by implementing a mitigation that involves a client-side call to a validation API. The fix is not perfect, as it still allows leaks and renders requests to arbitrary domains at times. However, it is a step in the right direction. The exact details of the validation process are unknown, as ChatGPT is not open source. The mitigation also remains limited to the web app, leaving mobile apps vulnerable. It is suggested that OpenAI limits the number of images rendered per response to further mitigate bypass trickeries. Overall, while not perfect, the mitigation is a positive development in improving security.
https://embracethered.com/blog/posts/2023/openai-data-exfiltration-first-mitigations-implemented/