Microsoft expands controls treating AI agents as insider threats
October 23: Dark Reading adds the August Defender preview, MDASH and MAI-Cyber-1-Flash, plus expert warnings that agents policing agents may raise risk.
Microsoft is extending Project Perception so enterprises can discover, identify, and constrain AI agents that hold company access.
Microsoft said on October 19, 2026 that enterprise AI agents should be treated as insider threats because prompt injection or task drift can change behavior mid-run. It is extending Project Perception, Agent 365, and Entra Agent ID, adding 35 local agent types for 60 total, including Claude Code, OpenAI Codex, and GitHub Copilot; TechCrunch instead names OpenClaw. Controls sit outside the model, with task-scoped Dynamic Permissions and Guardian Agents that can limit writes, require approval, isolate a runtime, or stop execution. Cyber Security News describes the work as planned rather than guaranteed prevention. Dark Reading says an August Defender preview used MDASH with MAI-Cyber-1-Flash and cites experts warning that agents policing agents may add risk.
- Discovery adds 35 local agent types, bringing the supported total to 60.
- Named types include Claude Code, OpenAI Codex, and GitHub Copilot.
- Entra Agent ID covers blueprints, principals, identities, and optional users.
- Dynamic Permissions start minimal, last for a task, and cannot exceed user rights.
- Rising risk can preserve reads, block writes, require approval, or halt execution.
- Project Perception was unveiled in July; a Defender preview followed in August.
Timeline
dated events in the reporting · oldest first- Oct 19, 2026Microsoft extends Project Perception and Entra Agent ID. · Microsoft Security Blog
- Oct 19, 2026Tenable maps Entra's four-object agent identity model. · Tenable Blog
- Oct 20, 2026Reports detail Dynamic Permissions and Guardian Agents. · SecurityWeek
- Oct 21, 2026Gallot says advanced models cannot be fully trusted. · CSO Online
- Oct 23, 2026Dark Reading describes August preview and expert warnings. · Dark Reading
What each outlet adds
beyond the shared facts- Microsoft Security BlogThirty-five new types reach 60; Entra covers sensor-only agents.
- Tenable BlogFour-object Entra model includes cross-tenant blueprints.
- TechCrunchNames OpenClaw and Nvidia's Open Agent Safety Platform.
- CSO OnlineProject Perception was unveiled in July.
- Dark ReadingAugust MDASH preview and expert warning on agent oversight.
Coverage
oldest first · one row per article · UTC- · 6d ago
Microsoft Security BlogFirst report
Microsoft extends Project Perception to secure AI agents as insiders
Discovery grows to 60 local agent types, and Entra Agent ID can cover agents found only through sensors.
- · 6d ago
Tenable Blog
Tenable One inventories Microsoft Entra Agent ID identities
Tenable says conventional IAM can miss Entra’s four-object agent model, including blueprints that span tenants.
- · 6d ago
Cyber Security News
Microsoft tells firms to treat AI agents as insider threats
Planned controls in Project Perception and Agent 365 would give agents separate identities and task-scoped, risk-adjusted permissions.
- · 5d ago
TechCrunch · AI
Microsoft launches Guardian Agents to monitor enterprise AI agents
Hayete Gallot says frontier models cannot be trusted because they are unpredictable, not malicious.
- · 5d ago
Security Affairs
Microsoft frames AI agents as insider threats needing live controls
Dynamic Permissions start at zero and last only for a task; Guardian Agents follow human playbooks.
- · 5d ago
Help Net Security
Microsoft announces Guardian Agents to constrain AI agent permissions
Discovery now covers 60 local agent types, and risk signals can revoke access or stop an agent.
- · 5d ago
SecurityWeek
Microsoft adds dynamic permissions and risk scores for AI agents
Project Perception now starts agents with minimal access and can isolate or halt high-risk runtimes.
- · 3d ago
CSO Online
Microsoft says AI agents should be treated as insider threats
Guardian Agents, part of Project Perception, are meant to watch other agents continuously and intervene on harmful actions.
- · 2d ago
Dark Reading
Microsoft expands Project Perception to contain anomalous AI agents
New capabilities discover agents, assign identity and permissions, score risk, and can cut off access.