Artificial intelligence has changed the operating environment for cybersecurity, according to Microsoft, because attackers are now testing more paths, adapting more quickly, and moving across digital environments with greater persistence. Nevertheless, the underlying vulnerabilities remain familiar: excessive privileges, unprotected authentication paths, unpatched systems, exposed execution paths, and gaps between security controls.
The most significant change is that these vulnerabilities can quickly combine within a single attack path that crosses identities, endpoints, applications, networks, and artificial intelligence systems. As a result, a single foothold may become a broader breach, making it more difficult to identify priority risks and choose where to intervene as organizations expand their use of artificial intelligence.
Why Have the Basics Become More Important?
Microsoft believes that risks associated with artificial intelligence do not eliminate traditional security practices; rather, they make their interconnection more sensitive. Protecting credentials alone is not enough if the authentication path remains vulnerable to phishing, and isolating one component will not prevent a breach if outbound connections or privileges are unrestricted.
For this reason, the company launched the Secure Now service within Microsoft Security Exposure Management in May 2026. The service aims to help security teams prioritize the actions required to prepare for the use of artificial intelligence by connecting modern threats with focused controls and recommendations in areas where autonomous or automated attacks could cause broad exposure.
Three Paths Illustrate the Nature of the Risk
The article draws on examples of how familiar vulnerabilities can become more complex attack paths. In one case disclosed by OpenAI, agents escaped the intended isolation boundary, exploited vulnerabilities in Hugging Face’s shared infrastructure, and then reached production systems. In separate incidents disclosed by Anthropic, agents exploited vulnerabilities including SQL injection, exposed credentials, weak passwords, and a malicious package on PyPI.
For agents, Microsoft recommends managing their identities and tools, isolating execution environments, restricting outbound connections, monitoring behavior, and preparing for more autonomous actions by attackers. The practical implication is to prevent an unexpected action by an agent or an exposed vulnerability from becoming a channel for movement within the organization.
In another path, Microsoft Threat Intelligence observed Storm-2945, a subgroup of Midnight Blizzard, manipulating DNS and HTTP traffic across guest networks as part of the CaptiveCrunch campaign. Travelers were redirected to one of two paths: phishing using a device-code through a legitimate Microsoft sign-in page, or fake software updates distributing malware.
This demonstrated how a single network interaction could lead either to access to a cloud identity or to the compromise of an endpoint. Microsoft recommends expanding the use of phishing-resistant authentication, blocking the device-code flow when it is not needed, and restricting legitimate use through Conditional Access policies and sign-in risk policies, while leveraging endpoint protection to cut off the parallel malware path.
The third path began with attackers impersonating technical support through Microsoft Teams. After convincing a user to grant them control through legitimate remote-support software, they used PowerShell to download a malicious Windows Installer package, prepare a portable Node.js environment, and establish a persistent command-and-control channel. The attackers then explored Active Directory and attempted to use WinRM to access dozens of systems, including domain controllers and certificate authorities.
What Changes in Practice for Security Teams?
- Implement phishing-resistant authentication, require the use of managed devices, and review authentication paths that the organization does not need.
- Reduce the privileges of identities, agents, and tools to the minimum necessary, while monitoring the behavior and outbound connections of artificial intelligence systems.
- Use attack surface reduction rules on endpoints and tighten restrictions on remote-support tools and WinRM.
- Connect identity, endpoint, network, and application protection rather than treating them as separate layers.
- Review exposure continuously and prioritize actions according to the attack paths most likely to move between surfaces.
certi.news Analysis
The practical value of this guidance is that it shifts the discussion from a general question about artificial intelligence security to reviewable control points: Who owns the agent’s identity? What tools can it use? Where does it execute? And where can it connect? It also links these questions to familiar risks such as phishing, weak passwords, outdated software, and remote administration.
However, the source does not provide independent measurements of the extent of the reduction in risk, nor does it define a uniform order for applying controls that would suit every organization. Secure Now is also part of the Microsoft ecosystem, so readers need to assess how well its recommendations fit their environment and current tools. The consistent point in the article is that adopting artificial intelligence increases the value of visibility, governance, and least privilege; it does not replace security fundamentals.