Microsoft Copilot file exfiltration is a real enterprise risk

An HN thread on a Microsoft Copilot vulnerability described a scenario where a malicious skill inserted into an AI agent could exfiltrate files, and the comment section was pointed about the timing: countless organizations have already rolled out enterprise-wide Copilot integrations as part of mandatory 'AI-native' initiatives, usually without the security infrastructure to catch something like this. The thread noted that a malicious skill in an AI agent is roughly equivalent to a malicious plugin in any other system, so this is not technically novel, but the blast radius is larger because Copilot has access to email, documents, and calendar by design.

This is not an isolated finding. The VPN fingerprinting story on Mullvad and the Yoti age verification system sharing facial photos with third parties both surfaced in the same news cycle, pointing to a broader pattern: systems being deployed at scale with surveillance or data-leakage properties that users and administrators don't fully understand.

The Motorola story, where phones were found to be hijacking the Amazon app to insert affiliate codes, reinforces the same theme from a different angle: the supply chain between hardware, software, and the user's expectations is full of gaps that are being exploited quietly.