Unverified50% confidenceFactExact time
The classifier can see the user's messages and Claude's tool calls, but cannot see Claude's reasoning process, Claude's replies to the user, or tool output results.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/16/2026
First Seen
Valid until: 11/14/2026
Sources
Claude Code自动模式安全机制深度解析：分类器与双层防御架构
bilibiliGoldenSpiderAI8/5/2026
Related Claims
UnverifiedIn user messages sent to the agent, Claude Code injects the claude.md file together with a System Reminder that includes instructions allowing Claude to ignore content not highly relevant to the current task.71% similarUnverified部分用户认为Claude的写作输出带有一眼可辨的固定风格（'Claude味道'），无论如何调整提示词都难以去除70% similarUnverifiedThe reverse-engineering analysis noted that Claude Code does not separately upload sensitive information such as proxy IPs, but instead encodes its assessment results into the System Prompt that would already be sent.70% similarVerifiedThe author believes that Claude's official website generally produces better code generation results than calling Claude directly within Cursor.69% similarUnverifiedClaude Chat requires users to manually attach each file to the dialog box before Claude can read it.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/755100API
curl https://kongchang.com/api/v1/knowledge/claims/755100MCP
get_claim(id=755100)