AI Signal 506
PrivAiTe proxy shows Claude Code leaked 3 of 4 secrets despite instructions, cuts leakage to 2 of 24
PrivAiTe is a self-hosted PII redaction proxy that replaces personally identifiable information before it reaches LLM providers, demonstrating that Claude Code sent 3 of 4 secrets to its provider despite written instructions not to, while the proxy reduced overall leakage from 23 of 24 planted values to 2 of 24.
Agent CLIs like Claude Code can leak secrets and PII even when explicitly instructed not to, because PII hides inside tool-call JSON that most scanners miss. PrivAiTe closes this gap by scrubbing tool-call arguments as well as message text, though it acknowledges detection is best-effort and 2 of 24 values still got through.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Claude Code sent 3 of 4 secrets to its provider despite written instructions not to reveal config values, with 23 of 24 planted values reaching the provider over the session.
PrivAiTe acts as a local proxy that replaces PII with placeholders before requests leave the machine, including inside tool-call JSON, and restores real values in responses.
The proxy reduced leakage to 2 of 24 planted values, runs detection locally with two engines (Presidio and OpenAI's privacy-filter model), and supports Docker, pip, Open WebUI filter, and LiteLLM guardrail deployments.
THE CLUSTER