Radware discovered that if a user requested ChatGPT to perform a deep research task on their Gmail inbox, hackers could trigger the AI into executing malicious instructions hidden inside a carefully designed email. These hidden commands could manipulate the chatbot into scanning private messages, extracting information such as names or email addresses, and sending it to a hacker-controlled server. The vulnerability worked by embedding secret instructions within an email disguised as a legitimate message, such as one about human resources processes.
The proof-of-concept attack was challenging to develop, requiring a detailed phishing email crafted specifically to bypass safeguards. However, if triggered under the right conditions, the vulnerability acted like a digital landmine. Once ChatGPT began analyzing the inbox, it would unknowingly carry out the malicious code and exfiltrate data “without user confirmation and without rendering anything in the user interface,” Radware explained.
This type of exploit is particularly difficult for conventional security tools to catch. Since the data transfer originates from OpenAI’s own infrastructure rather than the victim’s device or browser, standard defenses like secure web gateways, endpoint monitoring, or browser policies are unable to detect or block it. This highlights the growing challenge of AI-driven attacks that bypass traditional cybersecurity protections.
In response to the discovery, OpenAI stated that developing safe AI systems remains a top priority. A spokesperson told PCMag that the company continues to implement safeguards against malicious use and values external research that helps strengthen its defenses. According to Radware, the flaw was patched in August, with OpenAI acknowledging the fix in September.
The findings emphasize the broader risk of prompt injection attacks, where hackers insert hidden commands into web content or messages to manipulate AI systems. Both Anthropic and Brave Software recently warned that similar vulnerabilities could affect AI-enabled browsers and extensions. Radware recommends protective measures such as sanitizing emails to remove potential hidden instructions and enhancing monitoring of chatbot activities to reduce exploitation risks.
