Skip to content
Overseas4 min full readContains AI-generated content · human-reviewed

Microsoft extends Project Perception to internal agents, saying frontier models must be assumed untrusted

A Microsoft security executive said frontier models cannot be trusted regardless of origin, as the company extended Project Perception to agents running inside the enterprise.

  • Security & governance
  • Platforms & agents
A digital lock shown on a phone screen
Library photography for illustration, not event footage; depicted people and brands are unrelated to the events covered.

01What happened

Execution Containers is now generally available, sandboxing agents on Windows, macOS and Linux.

The company's CEO laid out the fuller frame on 10 October: assume every model is compromised, separate the model from its harness, log every meaningful action as tamper-proof human-readable evidence, and guarantee one authorised person can stop a task mid-run.

02By the numbers

3
platforms covered by Execution Containers
1
authorised person required to stop a task mid-run

03Background

The company's CEO put it more bluntly on 10 October: assume every AI model is compromised, separate the model from the harness that orchestrates it, and guarantee one authorised person can stop a task mid-run. Execution Containers' process containers already sandbox agents on Windows, macOS and Linux.

04Timeline

  1. 01
    October 10, 2026

    The company's CEO argues in public that every model should be assumed compromised

  2. 02
    October 19, 2026
    Event day

    Microsoft's security lead extends Project Perception to internal agents

Related items

A hand holding a padlock

OpenCode says an incident affected 1.4% of Console users

OpenCode said on 17 October that an incident the previous night, tied to a large-scale distillation attack, affected 1.4% of OpenCode Console users; it emailed those affected and is working with Anthropic, OpenAI and Google.

Read the record
A metal fastener clamping a stack of paper into one document

Anthropic hands its agent-skill format to a neutral foundation

Anthropic donated Agent Skills, its open standard, to the Linux Foundation's Agentic AI Foundation — including the name, logo, GitHub organisation and domain. More than 70 platforms have integrated it, with about 1M skills published and 280M installs.

Read the record

Want to apply these facts to your business? Bring the question Start AI diagnostic

This section is maintained by CodeLeo: topic and structure are ours, imagery comes from a licensed library, and the text runs through fact sheet → AI-generated draft → human review. The copy therefore contains AI-generated content. It is a public record of facts, states facts only, carries no commentary, and does not provide newsgathering or syndication services.

Company, product and brand names appear only to identify them. This section has no affiliation, agency, sponsorship or partnership with those companies, and nothing here recommends or endorses their business.