Skip to content
WeChat · Goumang Security Lab·Indexed by this site · Originally published· Tools →· Its place in the scan record →SelectedAttention40

veil local redaction proxy tested: real customer values in AI coding assistant requests are replaced with pseudonyms, and the request bodies received by the upstream contain no real values

「AI编程助手会把客户数据发给模型吗?veil本地脱敏代理实测」

AI summary

Jumang Security Lab empirically tested the local redaction proxy veil: real customer values in AI coding assistant requests are replaced with format-valid pseudonyms before being sent out, and the request bodies recorded by the fake upstream contain no real values. veil (PyPI package veil-proxy 0.3.2, alpha stage) learns real values from sources such as databases, fixtures, and git history, and uses HMAC to generate deterministic pseudonyms; the veil author states that the prompt cache hit rate for real Claude Code sessions is 74%–89%. The test showed that base64-encoded email addresses were redacted in their entirety, escaping was applied when restoring into shell commands, and WebFetch URLs retain pseudonyms to block leakage. The veil author's self-test states that on 1,000 labeled ai4privacy texts, pure regex let through 80.4% of names and contact information, while seeded veil let through 1.8%.

Why it was selected

The author used a fake upstream to intercept request bodies and empirically test veil's redaction and restoration effects; its records of usage boundaries such as Chinese names can serve as a reference for teams evaluating similar solutions.

Read originalmp.weixin.qq.com