Microsoft recently patched a critical vulnerability in its M365 Copilot AI platform that could allow attackers to retrieve sensitive data, including 2FA codes, from emails. The issue stems from the AI’s inability to distinguish between user instructions and malicious requests hidden in third-party content. This flaw is not unique to Microsoft, and other large language model providers are also struggling to secure this boundary, leading to the implementation of complicated guardrails to mitigate the consequences. Why it matters: this vulnerability highlights the significant security risks associated with AI models that can be exploited by attackers, making it a crucial concern in the broader AI landscape.