Unverified50% confidenceFactExact time
The untrustedContentHint: true annotation marks user-generated content or external data returned by a tool as untrusted, to defend against indirect prompt injection.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
WebMCP Security Guide: Three Hints That Make AI Agents Safer
youtubeChrome for Developers10/30/2026
Related Entities
Related Claims
Verified间接提示注入可将恶意指令藏入外部数据源（如网页、文档），待代理读取时自动触发攻击73% similarUnverifiedThe structural foundation for defending against prompt injection is explicit provenance marking and trust levels, rather than relying on instructions forbidding the model to follow web content.72% similarUnverifiedMost teams only guard user messages and assume retrieved content, file names, third-party API responses and error messages are safe, yet these are common entry points for indirect injection.71% similarUnverifiedThe case in this article is a typical indirect injection, which is more covert because the malicious text comes from a third-party data source trusted by the agent rather than from a directly interacting user.70% similarUnverifiedThe content of the model field in an API response payload can be filled in freely by the service provider and cannot be fully trusted; behavioral testing is a more reliable verification method.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1117378API
curl https://kongchang.com/api/v1/knowledge/claims/1117378MCP
get_claim(id=1117378)