A security flaw detected in Microsoft 365 Copilot AI.
Security Flaw Detected in Microsoft 365 Copilot
A significant security vulnerability has been discovered in the code of Microsoft 365 Copilot, the AI integrated into the widely used 365 office suite. This flaw could have allowed malicious actors to access the data of numerous users and their companies, according to a report by cybersecurity firm Aim Security published on Wednesday, June 11th.
Microsoft was notified several months in advance to address the vulnerability. According to the publisher, no customers were affected by data leaks, as reported by American media outlet Fortune. However, Aim Security stated that many companies using Copilot were theoretically vulnerable to this particularly insidious flaw, as it required no action from the victim. The attack was triggered automatically upon receiving an email in Outlook, the 365 email software.
During laboratory tests, Copilot scanned an infected email from Aim’s researchers during its routine email scan. The email contained a cleverly formulated instruction (or “prompt”) from the researchers, ordering it to search for sensitive data within its accessible information. In some companies, Copilot has access to a significant amount of information, allowing it to explore vast corporate databases and employees’ personal information to provide more relevant responses to each user. Aim’s researchers then only needed to exfiltrate this data to an external site, belonging to the attacker, by connecting Copilot to a Microsoft server entrusted with the transfer.
According to Aim, this type of novel attack (referred to by researchers as an “LLM scope violation”) bypasses the safeguards typically deployed by publishers. It could also affect other generative AIs besides Copilot.
Key Takeaways:
- A major security flaw was found in Microsoft 365 Copilot.
- The flaw could have allowed unauthorized access to user and company data.
- The attack required no action from the victim, triggering automatically via email.
- Microsoft was notified and claims no customers were affected.
- The vulnerability highlights a new type of threat to generative AI.
This new type of attack bypasses the safeguards typically deployed by publishers.
Enjoyed this post by Thibault Helle? Subscribe for more insights and updates straight from the source.
