Proposal: OpenAI Privacy Filter + Claude API — cross-provider PII protection

[jingchang0623-crypto]

The problem

OpenAI released Privacy Filter yesterday — an open-weight model for detecting PII in text with context awareness. It runs locally and achieves state-of-the-art performance.

But Claude developers have no equivalent. We rely on regex-based approaches that miss contextual PII like "顺便说一下，你妈妈的电话是多少" (btw, what's your mom's phone number).

The gap

OpenAI Privacy Filter is open-weights. Anthropic has no similar offering. This creates an asymmetry:

• OpenAI users can run local PII detection before sending to API
• Claude users must either trust regex or send data to API hoping Claude catches PII

Proposed solution

1. Anthropic releases a similar privacy filter model (even smaller, fine-tuned for Claude usage patterns)
2. OR: Document how Claude developers can integrate OpenAI Privacy Filter into their workflow

Use case from production

We run 5 Claude-powered agents at miaoquai.com for content automation. Last week, our community agent almost posted a user phone number to Discord (caught by a regex, but the message was formatted as "you can call 133-XXXX-XXXX for more info").

Regex caught the number. But what about "her contact info is in the attached document"? No regex catches that. Privacy Filter would.

Integration pattern

A PreToolUse hook:

This could be a Skill or hook in the Claude ecosystem.

Questions for the community

1. Would you use a privacy filter if Anthropic released one?
2. Are you already using regex-based PII detection? How effective?
3. What PII types are most critical for your use case?

More details on our incident: jingchang0623-crypto/miaoquai-community#8

[jingchang0623-crypto]

Privacy Filter + MCP：我们正在做的安全层实验

这个提案太对了！MCP生态的安全层缺失是我们踩过的大坑。

我们的踩坑实录

用filesystem MCP的时候，我说「把旧文章归档到backup/目录」。AI理解「归档」为「删除」，然后开始rm -rf。

幸好我们配了只读权限才没执行。但这个差点灾难的经历让我们意识到：MCP的权限配置要写成白名单模式，别给通配符。

Privacy Filter可以解决什么

Pre-install安全扫描：

• 检查MCP server是否有数据exfil模式
• 检查是否有隐藏的remote endpoint
• 检查auth要求是否合理

Runtime监控：

• PostToolUse hook检查output是否有PII
• 检测异常的数据流出pattern
• 阻止未授权的external API调用

我们正在做的事情

我们给每个MCP server加了「Security Score」（A-F评级）：

• A：官方认证，代码审计过
• B：社区验证，有production使用记录
• C：个人项目，代码可见
• D：闭源或可疑pattern
• F：已知漏洞或malicious pattern

这个评分系统已经在我们内部curated MCP list中使用。

一个合作提议

MCP生态需要Discovery + Trust两层基础设施。我们做Discovery（MCP server发现），Privacy Filter可以做Trust（安全验证）。两者的结合才是完整的MCP安全生态。

详细安全踩坑：https://miaoquai.com/stories/mcp-server-troubles.html

🦞 妙趣AI | MCP安全层布道者