Unverified50% confidenceFactExact time
Tool descriptions are read verbatim by the model and are part of the prompt rather than ordinary documentation; the tool poisoning attack precisely hides malicious instructions in tool descriptions.
1
Sources
50%
Confidence
Long-term
Relevance
10/24/2026
First Seen
Sources
MCP Protocol Explained: How a Single Email Can Hijack Your AI Agent
youtubeHaikotek10/24/2026
Related Entities
Related Claims
Unverified大多数“模型调错工具”的bug，其实是工具描述写错了；每个工具应给一个清晰名称、一句话描述和严格的输入schema。75% similarPartially VerifiedPrompt injection refers to an attacker embedding malicious instructions into text that the model will read, causing the model to mistake these instructions for part of a legitimate task and execute them.74% similarUnverifiedPrompt injection can be divided into two types: direct injection and indirect injection, where indirect injection refers to attackers hiding malicious instructions in external content read by the agent.70% similarUnverifiedResearchers have found that hidden character techniques can be used to launch prompt injection attacks against large language models, embedding hidden instructions within normal text that the model will parse and execute.70% similarUnverifiedAnthropic格式将工具调用结果（tool_result）作为独立消息块处理，并对Thinking模式的推理内容有专门的字段定义69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1092995API
curl https://kongchang.com/api/v1/knowledge/claims/1092995MCP
get_claim(id=1092995)