Skip to main content
SAICA-KG

Tools

Rebuff

1,464

LLM Prompt Injection Detector

DetectionPre-generationFully autonomousAt riskTier 4PromptEnvironment
Maturity tier details

Computed tier: 4(raw score 1.77). Effective tier: 4.

Facets:documented_in=0evaluated_on=0integration_surfaces=0.5paper_citations=0recent_commit=0stars=1.27

Source of truth:pipeline/maturity/score.py

Description

Rebuff is a prompt injection detection system designed to identify and filter malicious injection attempts in LLM prompts before processing. It analyzes incoming prompts to detect adversarial patterns that could compromise application security or manipulate model behavior.

Why it's included

Occupies detection × pre_generation × fully_autonomous with a four-stage ensemble for prompt-injection defense: heuristics, a dedicated classifier LLM, a vector store of prior attacks, and canary-token leak detection on the outbound side. Distinguished from last-layer (pure low-latency classifier) by trading latency for richer defense-in-depth, and from llm-guard by being injection-specialist rather than a general PII/toxicity suite. Useful anchor for the ensemble-detector style of pre-generation defense.

Addresses failure modes

View raw YAML on GitHub →