AI Security Gaps: How Attackers Exploit Trust Boundaries
Recent attacks show that enterprise AI share a common weakness: they accept external input without proper security boundaries.
In June, security researchers found serious in four popular AI . Microsoft 365 , , , and a supply- attack called - all had the same problem. They accepted external input without a trust boundary. This means could send instructions directly to the AI system.
The , called SearchLeak, let steal emails from a victim's mailbox. The victim only needed to click a link. had a of three that let a low- user become an admin and run code remotely. had a path traversal that gave full control. The - worm infected npm packages and stole credentials.
Security experts say these attacks are not about zero-day exploits. They are about poor configuration and missing security controls. Many companies deploy AI without proper governance. They give these access to sensitive data and systems. The fix is to audit AI , apply patches, and enforce least- access. Companies must treat AI systems as untrusted until proven secure.