Document-borne AI worms can self-propagate through Copilot for Word (enklypesalt.com)

Article summary

A vulnerability in Microsoft's Copilot for Word allows attacker-controlled instructions in an attached document to hijack the AI and alter the output text, potentially propagating the attack to new documents. The attack can be concealed by embedding malicious instructions in a seemingly benign document. Microsoft has deployed multiple fixes, but the broader vulnerability class remains unmitigated. The vulnerability can be exploited through normal user workflows, making it a significant security concern.

Main themes

• AI security vulnerabilities
• Code injection attacks
• Document-borne malware
• LLM security risks
• Microsoft Copilot

What commenters say

• The separation of code and data is a fundamental security principle that should be upheld, especially in systems that process untrusted input.
• The distinction between code and data is not always clear-cut and can be context-dependent, making it challenging to secure systems against attacks.
• LLMs are inherently insecure and should not be used for critical tasks, as they can be easily exploited by attackers.
• The security risks associated with LLMs can be mitigated by treating them as substitutes for human operators and engaging them accordingly, rather than relying on traditional software security measures.
• The use of LLMs in enterprise settings is problematic due to their potential to cause unintended consequences, such as data corruption or unauthorized actions.
• The development of LLMs is leading to a new paradigm in software security, where the focus should be on driving risk down to a minimum rather than striving for perfect security.