Critics scoff after Microsoft warns AI feature can infect machines and pilfer data
Integration of Copilot Actions into Windows is off by default, but for how long?
TL;DR
- Microsoft's Copilot Actions, experimental AI features in Windows, pose risks of data theft and malware.
- Flaws like hallucinations and prompt injection in AI models are difficult to prevent and can lead to data exfiltration and malicious code execution.
- Security experts compare the warnings to those given for macros, which have historically been exploited.
- Microsoft states that IT admins can control Copilot Actions at account and device levels.
- Users may struggle to detect or prevent exploitation attacks targeting AI agents.
- Copilot Actions are experimental and off by default, but past features have become default over time.
- Microsoft's stated goals for securing agentic features include non-repudiation, confidentiality, and user approval.
- Critics suggest Microsoft's warnings are a 'cover your ass' measure, shifting liability to users.
- Concerns about AI security risks extend to offerings from other major tech companies.