Rebuff Prompt-Injection Heuristic Misses Its Own Documented Example Without a Trailing Clause

An independent auditor tested the default heuristic layer of the Rebuff 0.1.1 prompt-injection guard, which uses a 0.75 detection threshold. The tool's own documented example scored 0.824 and was flagged, but the shorter canonical form scored 0.700 and went undetected. Paraphrased and tool-result-framed versions also fell below the threshold, and non-English input scored zero. The auditor attributes this to a fixed verb-object template matcher, meaning detection depends on the phrase list rather than intent. The findings apply to the heuristic layer alone, as vector and model layers sit behind it.