AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’

Collected Oct 19, 2026

Microsoft introduced a set of enterprise security features on Monday aimed at the risk posed by AI agents, and the headline item is Guardian Agents, a capability that watches over the agents a company runs rather than the outside attackers those agents might face. Guardian Agents sits inside Project Perception, a product Microsoft brought out in June. The framing from Microsoft's security leadership is blunt: frontier models, whichever lab produced them, are not trustworthy enough to be left unsupervised.

Project Perception was originally built around an external adversary. It shipped with three classes of agents modeled on security exercises: a red team that behaves like an attacker, a blue team that detects the red team and assesses the weaknesses it finds, and a green team that fixes them. That design assumed the dangerous AI would come from outside the organization. The new positioning inverts the assumption. The agents an enterprise has bought and deployed are themselves treated as potential threats, analogous to a rogue employee, the classic insider risk.

The argument has been made most forcefully by CEO Satya Nadella on his private blog earlier this month. He urged companies to stop extending trust to model providers, even when those providers offer safety and security assurances for their agentic products, and to run separate oversight systems that the AI labs do not control to watch and constrain the agents those labs' products spawn. His stated principle is that responsibility for what an intelligent system does on an organization's behalf cannot be outsourced.

Hayete Gallot, who leads Microsoft's security business, put it more directly when announcing Guardian Agents, writing that frontier models cannot be trusted regardless of where they come from. Her reasoning is not that the models are malicious or easy to compromise, but that they are unpredictable: they are optimized to complete tasks, not to follow organizational norms or exercise judgment. Combined with the broad access many agents receive to data and tools, that unpredictability is the exposure Microsoft is selling against.

Beyond Guardian Agents, the rollout includes an agent detection feature in Microsoft Defender that Microsoft says can identify up to 60 types of agents employees bring into the workplace, naming Claude Code, Codex and OpenClaw among them. There are also controls that restrict what data and systems agents can reach, and that adjust an agent's permissions dynamically according to the task at hand. Where an agent appears to be heading toward a risky decision, the system can require a human to approve before it proceeds.

There is an obvious tension in Microsoft making this pitch. The company is closely tied to OpenAI and Anthropic, has invested in both, counts OpenAI among its largest cloud customers, and runs its own model lab. Warning enterprises away from trusting labs' assurances cuts against some of those relationships. At the same time, Microsoft would prefer companies buy agent oversight from it, the cloud provider, rather than from the model vendors. Setting the commercial angle aside, the underlying logic holds up: the software that audits agent behavior should be independent of whoever supplies the agents, the same separation expected between an auditor and the organization being audited. This is my reading, not a claim in the announcement.

The broader industry is converging on a similar posture. Microsoft was an inaugural member of Nvidia's new Open Agent Safety Platform, which drew backing from more than 100 other technology companies. That platform isolates agents in secure environments and can monitor and halt their behavior at the hardware level, functioning in effect as a kill switch. The announcement follows Microsoft's launch of Surface Laptop Ultra PCs built for agent workloads, which run on Nvidia chips and ship with agent security and controls built in.

For developers and platform teams, the practical shift is that agent identity, permissions and behavior become first-class security concerns rather than a feature of whatever framework an agent was built on. Discovery matters first: an organization cannot govern agents it does not know exist, which is the point of cataloging up to 60 types in Defender. Scoped, task-based permissions and human approval gates are a move away from granting agents standing credentials to production systems. In practice, a likely trade-off is friction, since approval steps and permission churn slow autonomous workflows that were sold on running unattended. A second question is coverage and accuracy, since detection is only as good as the catalog of agent types it recognizes, and agents built in-house or on less common tooling may fall outside it.

Why it matters: Enterprises running coding assistants and other agents with access to sensitive systems are being pushed to treat those agents as insider risks and to buy oversight from a party other than the model vendor. Teams adopting Claude Code, Codex or OpenClaw should expect Defender-style discovery to surface them, and should plan for scoped permissions and human approval steps that change how autonomous their workflows can be. The competitive subtext is that the cloud layer, not the model lab, wants to own agent governance, which is likely to shape how agent security products are bundled and priced.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Microsoft wants enterprises to closely watch the agents that employees use with systems not controlled by AI labs like OpenAI and Anthropic.