Microsoft Expands Project Perception, Treating AI Agents as Insider Risks

Microsoft CEO Nadella announced on October 19 an expansion of Project Perception, formally putting the "model as insider risk" concept into practice, adding capabilities for continuously assessing agent risk, dynamically adjusting permissions, and containing threats. The same day, Microsoft issued a security advisory urging enterprises to treat AI Agents as potential "insider risks"—not because agents are inherently malicious, but because they can access company data, apps, and tools at high speed and act on them, with behavior that changes as instructions, data, and context shift.

Confirmed

• Nadella personally announced the Project Perception expansion on October 19, with a post on the Microsoft security blog citing industry reactions after several public AI security incidents in recent months
• Newly launched Guardian Agents: designed to monitor AI Agents used inside enterprises, part of the Project Perception framework announced in June
• Microsoft's security lead explicitly stated that frontier models cannot be blindly trusted or fully relied upon
• Identity verification firm Sumsub has also begun treating AI Agents as "insider threats," enforcing zero-trust controls in enterprise environments; VectorWire AI reported on this

Why it matters

• Agents hold permissions to call tools and access APIs and data; once hijacked or overstepping their bounds, their destructive power far exceeds traditional software, which static permission management struggles to cover
• Microsoft, as a platform provider, institutionalizing "don't trust agents" signals that the enterprise AI security paradigm is shifting from defending against external attacks to also guarding against internal (including AI-generated) risks, with zero trust becoming the main thread of agent governance

2026-10-19 ~ 2026-10-21 · 6 related posts

• [source] Microsoft expands Project Perception to treat AI models as insider risks with external security controls — satyanadella · 2026-10-19
• Microsoft warns AI agents can't be fully trusted and could act as insider security threats — brucemacv · 2026-10-20
• [source] Microsoft warns AI agents are like insider threats as 'frontier models can't be trusted' — emmanuelvivier · 2026-10-20
• Microsoft and Sumsub treat AI agents as insider threats with zero-trust controls — VectorWireAI · 2026-10-20
• [source] Microsoft Expands Project Perception to Secure AI Agents at Scale with Agent Identities and Risk Controls — Cole Medin · 2026-10-21

1 near-duplicate retellings: sanjaykalra