Skip to content
EventPast event

Microsoft builds trust-aware controls to address indirect prompt injection

2 reports2 reporting sourcesLatest report

Start here

AI overview

To reduce the risk of indirect prompt injection, Microsoft is introducing a trust-aware deterministic control layer for M365 Copilot, establishing a trust boundary between tool outputs and the main LLM and classifying content as trusted or untrusted based on deterministic signals (such as emails from outside the organization). Untrusted content is isolated and converted, through a capability-restricted path, into a structured representation that removes malicious content, retains useful information and attaches source labels; the main model receives only trusted original text and the converted structured representation. This layer supplements classifier filtering and system prompt hardening, but Microsoft states that it cannot prevent every indirect prompt injection. In a note published on October 27, 2026, Microsoft added that the protection initially focuses on high-risk content such as external communications, and that tenants need to enable Outlook sender settings and consult the CVE-2026-55145 configuration guide.

Generated by AI from reports · Updated 24 days ago

Progress timeline

Oct 27
  1. Microsoft Windows Developer Blog
    Microsoft builds trust-aware AI controls: handling untrusted content separately

    Microsoft is introducing a trust-aware deterministic control layer for M365 Copilot that handles trusted and untrusted content separately to reduce indirect prompt injection risk. Untrusted content is isolated and converted through a restricted path with source labels attached; the main model receives only trusted original text and the converted structured representation. The protection initially focuses on high-risk content such as external communications; tenants need to enable Outlook sender settings and consult the CVE-2026-55145 configuration guide.

Oct 15
Oct 13
  1. Microsoft Community Hub
    How Microsoft builds trust-aware controls for M365 Copilot, handling untrusted content differently

    Microsoft has added a deterministic enforcement layer at inference time to M365 Copilot, placing a trust boundary between tool outputs and the main LLM, and classifying content as trusted or untrusted based on deterministic signals, such as emails from outside the organization. Untrusted content is converted through an isolated, capability-limited path into a structured representation that removes malicious content, retains useful information, and attaches source markers. This layer supplements classifier filtering and system prompt hardening, but cannot stop every indirect prompt injection.

Related discussion

Total 7 items