Interesting how every feature built on top of LLMs seems to open a new attach vector. This time it’s OpenAI’s new ChatGPT memory feature, meant to personalize interactions, which was exploited to inject false memories & even exfiltrate user data.

Posted on 25 Sep 2024