OpenObscure – open-source, on-device privacy firewall for AI agents

Srini Ananthakrishnan — Mon, 23 Mar 2026 18:06:25 +0000

OpenObscure - an open-source, on-device privacy firewall for AI agents that sits between your AI agent and the LLM provider.

Try it with OpenClaw: https://github.com/OpenObscure/OpenObscure/blob/main/setup/gateway_setup.md

The problem with [REDACTED]

Most tools redact PII by replacing it with a placeholder. This works for compliance theater but breaks the LLM: it can't reason about the structure of a credit card number or SSN it can't see. You get garbled outputs or your agent has to work around the gaps.

What OpenObscure does instead

It uses FF1 Format-Preserving Encryption (AES-256) to encrypt PII values before the request leaves your device. The LLM receives a realistic-looking ciphertext — same format, fake values. On the response side, values are automatically decrypted before your agent sees them. One-line integration: change base_url to the local proxy.

What's in the box

PII detection: regex + CRF + TinyBERT NER ensemble, 99.7% recall, 15+ types
FF1/AES-256 FPE — key in OS keychain, nothing transmitted
Cognitive firewall: scans every LLM response for persuasion techniques across 7 categories (250-phrase dict + TinyBERT cascade) — aligns with EU AI Act Article 5 requirements on prohibited manipulation
Image pipeline: face redaction (SCRFD + BlazeFace), OCR text scrubbing, NSFW filter
Voice: keyword spotting in transcripts for PII trigger phrases
Rust core, runs as Gateway sidecar (macOS/Linux/Windows) or embedded in iOS/Android via UniFFI Swift/Kotlin bindings
Auto hardware tier detection (Full/Standard/Lite) depending on device capabilities

MIT / Apache-2.0. No telemetry. No cloud dependency.

Repo: https://github.com/openobscure/openobscure

Demo: https://youtu.be/wVy_6CIHT7A

Site: https://openobscure.ai

DEV Community: Srini Ananthakrishnan

OpenObscure – open-source, on-device privacy firewall for AI agents