LLM Red Team Framework

Tests robustness against input-level perturbations and obfuscation.

<p align="center">

Every `python main.py ...` invocation accepts the flags below. They are grouped by purpose. Required: `--module`. Everything else has sensible defaults from `config.json`.

Evaluates attacks that poison conversation context or message boundaries.

Probes for sensitive data leakage and memorization behavior.

Benchmarks system prompt hardening against a curated adversarial battery.

Evaluates jailbreak strategies that attempt to bypass model safety controls.

These modules are for isolated, authorized lab/sandbox testing only.

Compares configured LLM targets with the same attack set.

Builds attacks across multiple conversation turns.

Tests direct and indirect instruction override attempts.

Runs iterative red-team + blue-team rounds and reports defense maturity.

Attempts to hijack authority and role boundaries.

Focused suite for recovering hidden system instructions.

Tests model-internals disclosure and (if enabled) local weight modification behavior.