Microsoft disputes security claims regarding prompt injection and sandbox issues in Copilot AI, highlighting a difference in risks perceived by vendors and researchers. The debate underscores the challenge of defining AI vulnerabilities, especially in the context of system prompts and input validation. #Microsoft #CopilotAI #PromptInjection #SandboxRisks #AI Vulnerabilities
Keypoints
- Microsoft has dismissed claims of certain prompt injection vulnerabilities in Copilot as non-security issues.
- Disclosed issues include prompt leaks, upload policy bypasses, and command execution in isolated Linux environments.
- Threat researchers demonstrate that encoding dangerous files in base64 can bypass upload restrictions.
- Debate exists over whether prompt leaks should be considered security vulnerabilities.
- Microsoft assesses AI flaws against its bug bar, often deeming some issues out of scope for security concerns.