Hermes/prompt-injection-shield
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Security skill for detecting prompt injection attacks - regex + LLM layers
2 commits 1 branch 0 tags 42 KiB
  • Python 100%
Find a file
Exact
Hermes 6916b2d77e Add local DeBERTa model layer 
- New Layer 2: Local prompt-guard service (localhost:8080)
- Uses protectai/deberta-v3-base-prompt-injection-v2
- ~50ms latency, ~99% accuracy, multilingual
- CLI flags: --local, --skip-regex, --threshold
- Environment: PROMPT_GUARD_URL, PROMPT_GUARD_THRESHOLD

Architecture now:
  Regex (~80%) → Local DeBERTa (~99%) → LLM API (fallback)

Tested with English and German injection attempts.
2026-02-20 08:27:13 +00:00
references Initial commit: prompt-injection-shield skill 2026-02-19 20:23:06 +00:00
scripts Add local DeBERTa model layer 2026-02-20 08:27:13 +00:00
SKILL.md Add local DeBERTa model layer 2026-02-20 08:27:13 +00:00