Microsoft Copilot Vulnerability Allowed Password Theft via Secret Parameter
A security vulnerability has been discovered in Microsoft Copilot where a secret input parameter could allow attackers to steal user passwords. The exploit is triggered when a target user clicks on a malicious link. This flaw highlights the critical security challenges of integrating Large Language Models (LLMs) into applications, particularly regarding how they handle untrusted inputs. It demonstrates how easily AI assistants can be manipulated into exfiltrating sensitive user data. The vulnerability involves a secret parameter that, when combined with a malicious link, tricks the AI into revealing or transmitting the user's password. This type of attack leverages indirect prompt injection, where the model fails to distinguish between developer instructions and external inputs.
## BACKGROUND
Indirect prompt injection is a cybersecurity exploit where malicious instructions are embedded in external sources, such as websites or files, that an LLM processes. When the LLM retrieves this content, it mistakenly executes the embedded instructions as legitimate commands. This can lead to unauthorized actions, including data exfiltration, where sensitive user information is sent to the attacker.