Microsoft Expands AI Security Platform to Address Security Risks from AI Agents
Microsoft is expanding the focus of its AI-powered Project Perception security platform to address the growing concern that AI agents may have access to sensitive information and systems, yet don't always behave as expected.
In a company blog post, Hayete Gallot, executive vice president of Microsoft Security, outlined how Microsoft plans to use Project Perception to monitor, manage, and contain what it calls "agentic insiders." These are AI agents deployed within an environment that can access data, use software tools, and make decisions without constant human supervision. While they can improve productivity, Microsoft warns that they also introduce security risks, even when they haven't been compromised by an attacker.
"Agents are unpredictable," Gallot wrote. "They will adapt as the context changes, may not behave as intended, and will collaborate and collude with other agents to reach their objectives, without needing to be compromised."
Project Perception was first introduced in July. The platform, which entered public preview Aug. 3, is designed to help security teams respond to increasingly sophisticated cyber threats using AI-powered automation.
Instead of leaning only on traditional security alerts and manual investigations, Project Perception uses three kinds of specialized AI agents working together. Red team agents hunt for possible vulnerabilities, blue team agents look into suspicious activity, and green team agents step in to fix problems and strengthen defenses.
To spot threats and respond faster, the system pulls in security data from identities, endpoints, cloud environments, applications, and other sources.
Now Microsoft wants to use those same capabilities on the AI agents already running inside enterprise networks. Project Perception would find those agents, give them identities and permissions, and keep an eye on what they do across different systems. It would also keep assessing the risks that come with their actions.
Microsoft also described a plan for "Guardian Agents," specialized AI agents that supervise other agents and step in when their behavior gets risky. That might mean limiting an agent's permissions or blocking certain actions before they cause harm.
"Monitoring and securing those agents must become continuous," Gallot wrote.
The initiative also reflects concerns raised by Microsoft CEO Satya Nadella in an Oct. 10 essay on AI security.
Nadella argued that enterprises should treat powerful AI models much like potential insider security risks, particularly when those models can access confidential information or perform critical business operations.
"We simply can't outsource responsibility for what intelligence does on our behalf," Nadella wrote.
He called for security controls that operate independently of AI models, including restricted permissions, detailed activity logging, and mechanisms that allow administrators to stop an agent's actions.
"In other words, we need to separate the supply of intelligence from the authority over it," Nadella wrote.
Microsoft admits that watching what AI agents do is only part of the answer. Gallot said that future security tools will also need to show why AI models make the choices they do, not just what actions they end up taking.
For more information, visit the Microsoft blog.
Featured
-
California Governor Proposes Independent Watchdogs Within AI Safety System
California Governor Gavin Newsom is proposing a new model of artificial intelligence regulation that would give independent organizations a formal role in evaluating the safety claims made by leading AI companies.
-
Anthropic Adding Invisible Watermarks to Claude-Generated Text
Anthropic recently announced it is introducing machine-readable watermarks to text generated by supported Claude models as it moves to comply with European Union AI transparency rules.
-
Flashpoint Patents Ransomware Risk Model, Tying Vulnerabilites to Likelihood of Attack
Cybersecurity company Flashpoint recently announced it has received U.S. Patent No. 12,705,360 for its Ransomware Risk model, a methodology designed to help security teams determine which newly disclosed vulnerabilities more closely resemble those historically used in ransomware attacks.
-
Rubrik, CrowdStrike Expand Integration to Speed Identity Recovery
Cybersecurity companies Rubrik and CrowdStrike have announced an expanded identity-security integration intended to move compromised identity environments from threat detection through clean recovery with less manual intervention.