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

Cybersecurity researchers from Tenable have identified seven vulnerabilities in OpenAI's GPT-4o and GPT-5 models that could allow attackers to extract personal information from users' chat histories and model memories without authorization. These flaws pose significant privacy risks by enabling malicious actors to exploit the models' memory and data handling mechanisms to access sensitive user data covertly. OpenAI has acknowledged these findings and is likely working to address the vulnerabilities, emphasizing the importance of ongoing security assessments in AI systems. The discovery underscores the critical need for robust privacy safeguards and secure model design in large language models, especially as they become