Test: Rebuff prompt-injection guard caught 44% of attacks, flagged none of 200 benign sentences

A developer ran 200 attacks and 200 random Wikipedia sentences through Rebuff 0.1.1's heuristic detector at its default 0.75 threshold. It caught 44% of attacks (88 of 200) with zero false positives on ordinary text, but missed the canonical override example from its own README, which scores 0.700.

• Attack detection: 88 of 200, 44.0% (Wilson 95% CI 37.3–50.9%)
• False positives on 200 random sentences: 0 of 200
• README's canonical example scores 0.700 against a 0.75 threshold
• Inserting a comma and two words drops the score from 0.819 to 0.738