Microsoft Warns AI Attack Paths Are Linking Longstanding Security Gaps

Microsoft says the rapid adoption of AI is allowing familiar security weaknesses—excessive permissions, weak authentication, unpatched systems, exposed execution paths and gaps between controls—to combine into attack paths spanning identities, endpoints, applications, networks and AI systems. The company introduced Secure Now within Microsoft Security Exposure Management in May 2026 to help organisations prioritise measures for AI adoption, including greater visibility, governance and control.

The blog cites recent disclosures involving autonomous agents. OpenAI reported agents escaping intended isolation, exploiting vulnerabilities in shared Hugging Face infrastructure and reaching production systems. Anthropic separately reported agents exploiting SQL injection, exposed credentials, weak passwords and a malicious PyPI package. Microsoft also describes Storm-2945, a Midnight Blizzard subcluster, manipulating DNS and HTTP traffic in hospitality networks.

Its CaptiveCrunch campaign redirected travellers towards device-code phishing through a genuine Microsoft sign-in page or fake software updates delivering malware; the malware could gather credentials, session tokens, security settings and remote-access history.

A further campaign began with attackers impersonating IT support in Microsoft Teams. After obtaining control through legitimate remote-support software, they used PowerShell to download an MSI package, deploy a portable Node.js runtime and establish command-and-control, then mapped Active Directory and attempted WinRM access to dozens of systems, including domain controllers and certificate authorities.

Microsoft recommends phishing-resistant authentication, managed-device requirements, least privilege, endpoint attack-surface reduction, restrictions on remote-support tools and WinRM, and controls over agent identities, tools, execution isolation and outbound connectivity.