Microsoft Cautions That AI Agents Might Mirror Insider Threats Due to ‘Frontier Models’ Trust Issues
On Monday, Microsoft announced a new suite of security features aiming to protect enterprises from the growing threats linked to AI agents. This rollout aligns with Microsoft’s ongoing alerts regarding the potential dangers posed by AI technologies and their agents.
A key feature introduced is Guardian Agents, designed to oversee the AI agents used within organizations. Guardian Agents is part of Project Perception, a product that was first launched in June.
Project Perception was Microsoft’s initiative in response to advanced hacking frameworks such as Anthropic’s Mythos and Fable, as well as the security vulnerability scanning tool from OpenAI named Daybreak. Initially, it featured three categories of agents: a “red team” that simulates hacker behavior, a “blue team” that identifies vulnerabilities indicated by the red team, and a “green team” dedicated to remediation efforts.
The project was based on the assumption that AI hackers would generally emerge from external threats. However, Microsoft CEO Satya Nadella has recently pointed out that the more significant threat might stem from the AI agents actively used by organizations themselves, similar to a rogue employee, often described as an “insider threat.”
Earlier this month, in a private blog entry, Nadella cautioned organizations to remain skeptical about the trustworthiness of AI model providers, even when these companies offer guarantees pertaining to the safety and security of their agent products. He advocates for establishing independent systems, separate from AI labs, to oversee and manage the agents activated by these labs’ offerings.
“We are deploying these sophisticated agentic systems and models that have access to our most sensitive data, granting them the authority to perform mission-critical actions on our behalf!” he emphasized. “We cannot simply relinquish responsibility for the actions of intelligence acting on our behalf. A model provider’s assurances do not relieve us of that duty. … Essentially, we must decouple the supply of intelligence from the power to control it.”
At the announcement of Guardian Agents, Microsoft’s security division chief Hayete Gallot was particularly forthright.
“Regrettably, frontier models cannot be trusted, regardless of their origin,” she remarked, clarifying that this is not due to malicious intent or ease of exploitation, but rather their unpredictable nature. These models are engineered to perform tasks, not to conform to “organizational norms or sound judgment.”
Furthermore, they often have unrestricted access to sensitive information and tools, which poses a risk of substantial damage.
Microsoft’s strategy is both surprising and, arguably, self-serving. The unexpected nature stems from Microsoft’s close connections with OpenAI and Anthropic, both of which it has invested in. OpenAI is a significant customer of Microsoft’s cloud services, and Microsoft also operates an AI lab that develops its models. Consequently, Microsoft is promoting skepticism towards some of its own business partnerships.
The self-serving element is that Microsoft aims for organizations to acquire agent security solutions directly from it, the cloud provider, rather than from Anthropic or OpenAI. Despite potential conflicts of interest, the rationale is sound. Software that monitors agent behaviors should be independent of the suppliers of those agents, just as auditors must remain unbiased toward the organizations they assess.
Or, as Nadella succinctly puts it: “We need to separate the supply of intelligence from the authority over it.”
Microsoft is seizing this business opportunity through various means. Beyond Guardian Agents, it has rolled out an agent detection feature within its Microsoft Defender product that can identify up to 60 different types of agents that employees might bring into the workplace, including Claude Code, Codex, and OpenClaw. Additionally, it has integrated features that limit agents’ access to data and systems and dynamically adjust their permissions based on tasks, requiring human oversight whenever an agent appears to be making questionable decisions.
The view of agents as unreliable insiders is gaining traction across the industry as well.
Microsoft was among the early adopters of Nvidia’s new Open Agent Safety Platform, which has received backing from over 100 other tech companies. This platform confines agents in secure environments and can monitor and terminate operations through hardware controls, functioning akin to a kill switch. This follows Microsoft’s launch of new Surface Laptop Ultra PCs tailored for agents, powered by Nvidia chips and equipped with built-in agent security and controls.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.


