Autonomous software is increasingly embedded in corporate operations, and Microsoft now argues the biggest risk may come from within. The company warns that AI agents deployed on company budgets could behave like an employee gone rogue if left unsupervised, and has announced a new set of defenses to counter that.

According to Microsoft, security commitments from model developers alone are not enough. Systems with access to sensitive data and critical permissions can cause serious damage to organizations when left unmonitored.

The trust question: "No matter the origin..."

2

Microsoft Security Unit leader Hayete Gallot puts it bluntly: "No matter the origin, the most advanced frontier models cannot be trusted." CEO Satya Nadella takes the same line, arguing that model makers should not be trusted unconditionally and that the supply of intelligence must be separated from control over that intelligence.

The rationale behind this caution is not malice but nature: these systems are programmed to complete their tasks rather than follow corporate rules, and can behave unpredictably.

Project Perception expands

Project Perception, announced in June to scan for security flaws in powerful systems from Anthropic and OpenAI, has been expanded with new components. The standout addition is "Guardian Agents," which monitor the AI agents a company uses in real time. The industry's defensive focus is shifting from externally sourced AI attacks to agents operating inside the organization.

Independent auditing logic and the hardware side

2

As the largest investor and cloud partner of both OpenAI and Anthropic, Microsoft's stance could read at first glance as distrust of its own partners. The company's underlying principle is different, however: the labs that build agents and the security layer that audits them should be independent of each other. That approach also reinforces Microsoft's position in the security market.

Concrete steps have followed. Microsoft Defender can now detect 60 different agent types, including Claude Code and Codex that employees bring into the workplace. Protocols have also been added that dynamically narrow agents' permissions and require human approval for risky decisions.

The message is resonating across the industry. Under NVIDIA's leadership, more than 100 companies are backing the "Open Agent Safety Platform," which is working on hardware-based emergency stop systems that isolate agents, while Microsoft is accompanying the shift with a new Surface lineup featuring built-in security controls.

The industry's defensive focus is shifting from externally sourced AI attacks to agents operating inside the organization.