Static analyzers flag emoji-zwj-sequence as Prompt Injection #259

Version: v2.3.11, repo HEAD c2d09df, static mode (--no-llm), running via the command line.

Observation: scan such as SKILL.md which contains emoji

Example:

1. Part of the content in SKILL.md: 🧑‍⚖️
2. The unicode sequence is: ['U+1F9D1', 'U+200D', 'U+2696', 'U+FE0F']
3. The scan result: { "id": "P2", "category": "Prompt Injection", "pattern": "Hidden Instructions", "severity": "HIGH", "confidence": 0.6, "location": { "file": "SKILL.md", "start_line": 147, "end_line": null }, "finding": **"\u200d"**}

Analyze:

1. the SKILL.md will match pattern (r"[\u200b\u200c\u200d\u2060\ufeff]", 0.6) in P2_PATTERNS
2. the matched \u200d which is part of emoji-zwj-sequence, (1F9D1 200D 2696 FE0F ; RGI_Emoji_ZWJ_Sequence ; judge) see detail in https://unicode.org/Public/emoji/16.0/emoji-zwj-sequences.txt, which is false positive