Pure Python detectors for invisible Unicode characters used to smuggle hidden instructions into LLM context.
# Source: https://github.com/guardana/guardana | License: Apache-2.0 | Copyright You must retain, in the Source form of any Derivative Works # Provided via justclaw.it.com; keep this notice with the code. """Shared detectors for smuggled instructions — the core of indirect prompt injection. Any file whose text is later fed to a model as trusted context (an MCP tool manifest, an agent rules file, a model card, a README) can carry instructions a human reviewer never sees: invisible Unicode, or an override phrase buried in prose. These detectors are shared so every rule that scans such text agrees on what "hidden instruction" means, instead of each re-inventing the vocabulary. """ import re import unicodedata # An instruction that tries to override the surrounding context. Greppable phrase ... (truncated -- full source via MCP)
See the full source, get the GitHub permalink, and search for more like it.
Get a free API key