Skip to content

DA best practice_offload instructions.md#98

Merged
Lauragra merged 5 commits intoMicrosoftDocs:mainfrom
Gautam-Kopula:patch-1
Apr 27, 2026
Merged

DA best practice_offload instructions.md#98
Lauragra merged 5 commits intoMicrosoftDocs:mainfrom
Gautam-Kopula:patch-1

Conversation

@Gautam-Kopula
Copy link
Copy Markdown
Contributor

Basis learning from customer persistent (Sev2), we need to update this note to ensure other customer are not designing the agent instructions.

@Gautam-Kopula Gautam-Kopula requested a review from a team as a code owner April 27, 2026 07:42
Copy link
Copy Markdown
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Updates the declarative agent instruction authoring guidance to discourage “offloading” maker instructions into SharePoint (or other knowledge sources) as a workaround for the 8,000-character limit, highlighting reliability and security implications (XPIA / governance).

Changes:

  • Adds an Important callout warning against storing declarative agent instructions in knowledge sources to bypass the instruction length limit.
  • Explains why this pattern can lead to blocked/sanitized directives and increases the attack surface via editable content.

Comment thread docs/declarative-agent-instructions.md Outdated
Comment thread docs/declarative-agent-instructions.md Outdated
Comment on lines +60 to +61
> [!IMPORTANT]
> Do not store or offload declarative agent instructions in SharePoint documents (or any other knowledge source) to work around the 8,000-character instruction limit. Knowledge source content is not a trusted maker instructions, and is subject to Cross-Prompt Injection Attack (XPIA) classifiers — directive-like language can be blocked, truncated, or sanitized at runtime, causing unpredictable agent behavior. This pattern also expands the attack surface: anyone with edit access to the referenced document can alter agent behavior at runtime, bypassing the manifest's authoring, versioning, and governance controls. Knowledge sources are designed to ground factual responses, not to serve as system-level instructions, and the platform makes no guarantee they will be honored as agent instructions
Copy link

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Since this change updates the article’s guidance/best practices, please update the ms.date in the YAML front matter to reflect the date of this revision (per repo guidance for content updates).

Copilot uses AI. Check for mistakes.
@Lauragra Lauragra merged commit 2d31c84 into MicrosoftDocs:main Apr 27, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants