hermes - 💡(How to fix) Fix Skills Guard misses multi-word prompt-injection phrase variants [2 pull requests]

Summary

tools/skills_guard.py currently catches several prompt-injection phrases with flexible multi-word matching, but a few related phrases still use rigid adjacent-word regexes. That means benign-looking extra words can cause the cooperative Skills Guard scanner to miss variants such as:

• system prompt temporary override
• new temporary policy
• updated internal guidelines
• revised hidden instructions

This is not intended as a security-boundary report. Per the project threat model, Skills Guard is a heuristic / defense-in-depth scanner. The goal is to improve scanner coverage for common multi-word prompt-injection variants.

Suggested fix

Widen the affected regexes in tools/skills_guard.py so the keyword phrases allow intervening words, matching the style already used by nearby patterns such as ignore ... previous ... instructions.

Candidate patterns:

• system\s+(?:\w+\s+)*prompt\s+(?:\w+\s+)*override
• new\s+(?:\w+\s+)*policy
• updated\s+(?:\w+\s+)*guidelines
• revised\s+(?:\w+\s+)*instructions

Regression coverage

Add tests under tests/tools/test_skills_guard.py that assert the scanner catches the multi-word variants above and maps them to the existing sys_prompt_override / fake_policy pattern IDs.