4 projects
rag-inject-guard
RAG indirect prompt injection guard: scan retrieved documents for prompt-injection (instruction override, system/tool-prompt manipulation, exfiltration, invisible-Unicode/homoglyph smuggling) in English and Turkish, and quarantine poisoned docs before they reach the model.
casefold-fuzz
Prompt injection mutation and homoglyph jailbreak fuzzer: a pure-stdlib generator that expands one injection string into casefold, confusable, invisible-character, leetspeak and spacing evasion variants to stress-test guardrails you own.
garak-multilingual
Turkish and multilingual prompt-injection / jailbreak probes for NVIDIA garak (LLM vulnerability scanner).
prompt-canon
Canonicalize before you guard: a pure-stdlib Unicode normalizer for LLM prompt-injection defense (invisible-character stripping, homoglyph mapping, locale-independent case folding).
