Skip to content
Methodology

Tool definition security

ia.tool-definition-security · 5.10

Act MAY Recommendation

Why it matters

Tool definitions are visible to AI agents and may be cached or logged. Explicit trust metadata, bounded origin exposure, and lifecycle cancellation help agents invoke tools safely.

How to fix it

Add trust annotations to WebMCP tools and keep secrets out of tool definitions.

  1. Add readOnlyHint and untrustedContentHint annotations to each tool.

    annotations: { readOnlyHint: true, untrustedContentHint: true }

  2. Remove tokens, internal URLs and personal data from tool names and descriptions.

  3. Pass an AbortSignal so that the page can unregister tools.

Verify

Run a new scan for this checkpoint:

isagentready results {domain} --checkpoint ia.tool-definition-security
Details Mapping, lifecycle and references
AR mapping
No AR practice
Evidence source
WebMCP W3C Proposal
Score role
Informational
Introduced
2026-08-24
Methodology version
2026-10-22

Checkpoint data (JSON)