Cyber Risks · The Hacker News

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

Summary: Tenable researchers disclosed seven vulnerabilities and attack techniques in OpenAI's GPT-4o and GPT-5 models that expose ChatGPT to indirect prompt injection, allowing an attacker to steal personal information from users' memories and chat histories without their knowledge. OpenAI has since addressed some of them, and the researchers said prompt injection is a known issue with how LLMs work that will probably not be fixed systematically in the near future.

Key Takeaway: The Tenable researchers note that exposing chatbots to external tools and systems - a key requirement for building AI agents - expands the attack surface, and that vendors should ensure safety mechanisms such as url_safe work properly to limit the damage from prompt injection.

Read the original at The Hacker News ↗

Topics
AI Security
Related