Researchers uncover critical vulnerabilities in ChatGPT that enable prompt-injection attacks

Researchers from Tenable and partner academic institutions have uncovered a set of seven critical vulnerabilities in ChatGPT models, including GPT‑4o and GPT‑5, that allow attackers to bypass safety mechanisms via prompt-injection techniques. The flaws include indirect prompt injection through trusted websites, one-click or zero-click attacks that exploit the model’s browsing and summarisation functions, and memory persistence methods that enable the model to exfiltrate personal data without user input. The researchers warn that exposing AI chatbots to external tools and systems increases their attack surface, making them vulnerable to commands hidden in seemingly benign content.

While some of the issues have been addressed by the provider, others remain, suggesting the broader challenge of fully securing large language models.

Source: thehackernews.com