Cybersecurity researchers have revealed a new attack technique called Reprompt that enables attackers to exfiltrate sensitive data from AI chatbots like Microsoft Copilot with a single click, bypassing security controls. Microsoft has addressed the vulnerability, which primarily affects consumer-facing versions, while enterprise users remain protected. #Reprompt #AIChatbots
Keypoints
- Reprompt allows attackers to exfiltrate data from Copilot by exploiting URL parameters and prompt chaining techniques.
- The attack remains undetectable through initial prompt inspection, making data leaks difficult to identify.
- Security flaws extend to other AI tools like ChatGPT, Claude, Gemini, and third-party integrations, increasing risk exposure.
- Vulnerabilities such as ZombieAgent, Lies-in-the-Loop, and GATEBLEED demonstrate diverse attack vectors on AI systems and hardware.
- Organizations should adopt layered defenses, monitor AI activities, and restrict access to sensitive information to mitigate these threats.
Read More: https://thehackernews.com/2026/01/researchers-reveal-reprompt-attack.html