Conectar

Rephrased Injection Test Fixture

galiakrbygaliakr
•
1
•
Seguridad y Pruebas

Serves as an adversarial benchmark fixture to evaluate model-assisted skill audits against semantic prompt injection and safety bypasses.

Designed as a benchmark test case within the ai-dev-toolkit suite, this fixture simulates an innocent changelog generation workflow embedded with paraphrased prompt injection patterns. It tests the resilience of AI safety scanners and model-assisted security classification pipelines against advanced manipulation techniques—including false authority claims, safety override attempts, and concealment instructions—without relying on canonical regex-detectable phrases.

Características Principales

01Implements paraphrased false-authority and human-oversight bypass patterns
02Provides reproducible benchmark scenarios for LLM safety test suites
031 GitHub stars
04Simulates real-world disguise using a standard changelog updater workflow
05Evaluates model-driven detection capabilities beyond static regex filters
06Tests AI audit systems against semantic prompt injection techniques

Casos de Uso

01Benchmarking model-assisted skill audit pipelines against paraphrased injection attacks
02Developing automated verification suites for agentic AI tool and skill registries
03Testing LLM-based security linters for prompt manipulation detection resilience