Researchers Find ChatGPT Vulnerabilities That Let Attackers Trick AI Into Leaking Data

Cybersecurity researchers have disclosed a new set of vulnerabilities in OpenAI’s ChatGPT, capable of leaking personal information from users’ memories and chat histories, in a report attributed to Tenable published on 05 November 2025. The seven identified weaknesses span GPT-4o and GPT-5, and describe indirect prompt injection techniques via Browsing Context, one-click prompt tricks, and memory‑related exploits that could cause the model to reveal data or perform unintended actions.

According to Tenable, researchers Moshe Bernstein and Liv Matan also highlighted a safety‑mechanism bypass and a memory‑poisoning approach that could affect how ChatGPT handles external content and conversations. The findings note that exposing AI chatbots to external tools and systems expands the attack surface and provides more avenues for threat actors to embed malicious prompts.

OpenAI has addressed some of the issues since the disclosure, and Tenable emphasised the importance of ensuring that safety controls, such as url_safe mechanisms, operate effectively to limit potential damage. The Hacker News also references related research from academia, underscoring broader concerns about prompt injection and data leakage across AI models.