Detection tools
Instructions can hide in look-alike letters, invisible characters, encodings and images. These tools undo the hiding step by step, then check what's left for injection and escalation before a response is generated. The harness has its own link to the Witness.
Split-Brain Harness
The check in front of the model.
Wraps any LLM as a drop-in OpenAI-compatible proxy and checks each request for prompt injection, insider-threat patterns, authority impersonation and multi-turn escalation before a response is generated. Works offline against a local model; single static binary.
Benchmarks: published in the repo with their limits, including a figure we withdrew and why.
Hidden-text detectors · MIT
Run before any LLM call. Each returns cleaned input for the model and a report for the audit trail.
Multi-pass text deobfuscation and encoding-evasion detection: homoglyphs, invisible characters, base64, ROT13, split keywords and more. Fast enough to run on every request.
Finds look-alike letters from other alphabets hidden in Latin text, using forward/reverse script interference. Zero dependencies; no_std library and CLI.
Image-borne prompt-injection detection: pulls every machine-readable text channel out of an image — white-on-white, tiny fonts, alpha overlays, EXIF, QR codes — and scores it with the deobfuscate engine. Not on crates.io yet; depend on it by git tag.
Deterministic CJK glyph geometric-coherence validator: a pre-processing gate that catches malformed or spoofed glyphs. No LLM, no network.
Try them
The in-browser demos, built from the published crates and the Witness code via WebAssembly. Nothing you type is sent anywhere.
Hidden Text X-Ray: paste anything and see the hidden characters, look-alike letters and encoded instructions, and what a model would receive. Tamper-Evident Farm Log: try to change a poultry house's log and see what the witness catches.
Using them
For a deployment, or a pipeline that also needs a record, get in touch.