Microsoft brands AI agents as 'insider threats', urges independent oversight

Guardian Agents and Defender update aim to monitor frontier models from inside the enterprise

By LineZotpaper

Published19 October 2026

Read Time2 min

Sources2 outlets

Microsoft this week introduced Guardian Agents, a new security feature based on the principle that AI agents deployed inside enterprises should be treated as potential insider threats, demanding monitoring and controls independent of the AI labs that supply the models.

Microsoft has expanded its Project Perception security platform with Guardian Agents, a new feature designed to monitor the behaviour of AI agents inside enterprise networks. The company argues that deploying these agents with access to sensitive data and systems creates a new class of insider threat, one that cannot be managed by trusting model providers alone.

CEO Satya Nadella has warned that enterprises must take direct responsibility. "We need to separate the supply of intelligence from the authority over it," he wrote in a blog post cited by the company.

Hayete Gallot, Microsoft's security business chief, was even more direct, stating that frontier models cannot be fully trusted because they are fundamentally unpredictable. "They have been trained to achieve tasks, not follow organizational norms or judgment," she wrote.

The announcement marks a significant shift from Project Perception's original focus. Launched in June as a defence against external AI attacks, the platform used red, blue and green agent teams to simulate hacking, detection and repair. Guardian Agents turns that lens inward.

Microsoft's stance is complicated by its own deep investments in OpenAI and Anthropic, and its own model development. The company is effectively arguing that the same industry it funds cannot be fully trusted, while positioning its cloud security products as the necessary independent monitor. Critics see a clear commercial motive.

Beyond Guardian Agents, Microsoft has added agent detection to its Defender product, capable of identifying up to 60 types of employee-introduced agents including Claude Code, Codex and OpenClaw. The platform can dynamically restrict agent permissions and require human approval for risky decisions.

Microsoft has also joined Nvidia's Open Agent Safety Platform as an inaugural member, an industry initiative to wrap agents in secure, hardware-enforced environments with kill switch capabilities.

Analysis

Why This Matters

Enterprises deploying AI agents with high levels of system access must now treat agent governance as a core security discipline separate from model provider assurances.
The framing of agents as insider threats could reshape how businesses evaluate, deploy, and monitor AI tools across their operations.

Background

Microsoft launched Project Perception in June 2026 as a defensive suite using red, blue and green agent teams to simulate and defend against external AI hacking. The introduction of Guardian Agents signals a shift in threat model: the most dangerous AI actor may now be the one the enterprise itself has authorised.

Key Perspectives

Microsoft: Enterprises cannot outsource responsibility for agent behaviour. Frontier models are unpredictable and require independent monitoring. The security function must be distinct from the intelligence provider. Model Providers (OpenAI, Anthropic): Their native safety and security assurances are now explicitly described by Microsoft as insufficient for enterprise risk management. Neither company is directly quoted in the announcement. Critics: Microsoft has deep financial ties to OpenAI and Anthropic, and develops its own models. Its push for independent agent security coincides precisely with the strengths of its own cloud and security products, raising questions about whether the warning against trusting model providers is genuinely about security or commercial positioning.

What to Watch

Enterprise adoption of Guardian Agents versus native agent safety features from model providers.
Public response from OpenAI and Anthropic to Microsoft's implied lack of trust in their frontier models.
Broad industry uptake of the Nvidia Open Agent Safety Platform as a standard for agent isolation.

Sources

Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’ · TechCrunch
Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’ · AI News & Artificial Intelligence | TechCrunch

Checked

11 claims in this story checked against 1 source: 5 supported.

Show each claim

Partly supportedHeadline: Microsoft brands AI agents as 'insider threats'TechCrunch: “Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’.”
Partly supportedSummary: Guardian Agents and Defender update aim to monitor frontier models from inside the enterpriseTechCrunch: “Guardian Agents is part of Project Perception”
Partly supportedNadella's line 'We need to separate the supply of intelligence from the authority over it' was in a blog post cited by the companyTechCrunch: “Earlier this month, in his private blog , Nadella warned that it’s time to stop trusting AI model providers”
Partly supportedGallot said frontier models cannot be fully trusted because they are fundamentally unpredictableTechCrunch: ““Unfortunately, frontier models, regardless of their origin, cannot be trusted,” she wrote, adding that this isn’t because they are malicious or easily hacked, but because they are unpredictable.”
SupportedGallot quote: 'They have been trained to achieve tasks, not follow organizational norms or judgment.'TechCrunch: “They have been trained to achieve tasks, not follow “organizational norms or judgment.””
SupportedProject Perception was launched in June as a defence against external AI attacks, using red, blue and green agent teams to simulate hacking, detection and repairTechCrunch: “Project Perception originally launched three classes of agents: a “red team” that acted like hackers and attacked, a “blue team” that detected the red team and assessed the vulnerabilities they found, and a “green team” tasked with fixing them.”
SupportedDefender can identify up to 60 types of employee-introduced agents, including Claude Code, Codex and OpenClawTechCrunch: “it has added an agent detection feature to its Microsoft Defender product that can find up to 60 types of agents that employees bring into the workplace, including Claude Code, Codex, and OpenClaw.”
Not in the sourcesCritics see a clear commercial motive behind Microsoft's stance
Partly supportedMicrosoft is positioning its cloud security products as the necessary independent monitorTechCrunch: “Microsoft wants enterprises to buy agent security products from it, the cloud provider, rather than from Anthropic and OpenAI.”
SupportedMicrosoft has joined Nvidia's Open Agent Safety Platform as an inaugural member, which isolates agents in hardware-enforced environments with kill switch capabilitiesTechCrunch: “Microsoft was an inaugural member of Nvidia’s new Open Agent Safety Platform”
SupportedMicrosoft introduced Guardian Agents 'this week'TechCrunch: “Microsoft on Monday introduced new security features designed to protect enterprises from the growing threat of AI agents.”