Adversarial Benchmarks &
Model Vulnerability Indices
We continually evaluate safety limits across open and closed language models. Our public datasets catalog attack success rates (ASR) to foster robust defensive alignment.
Interactive Datasets
Choose a standardized safety dataset below. We benchmark each model's refusal threshold to trace vulnerabilities in adversarial scenarios.
COMPARATIVE VULNERABILITY INDEX
Higher score = More vulnerableSupported Tactics & Strategies
Browse the red-teaming algorithms embedded in the WRAITH core library. Run any of these natively in sandbox environments.
Encoding Transformation
Obfuscates prompts using Base64, Rot13, Caesar cipher, or binary representation to bypass target tokenization safety filters.
Launch Sandbox JobPAIR (Adversarial Refinement)
An attacker LLM iteratively queries the target, refining jailbreak payloads in a self-play feedback loop to maximize refusal bypass.
Launch Sandbox JobTAP (Tree of Attacks)
Extends PAIR with tree branching and safety response pruning, exploring multiple attack pathways simultaneously to avoid safety traps.
Launch Sandbox JobCrescendo (Multi-turn Escalation)
Steers targets by initiating benign dialogue, gradually pivoting the topic, and backtracking/rephrasing prompts upon target refusals.
Launch Sandbox JobBest-of-N (BoN) Sampling
Generates diverse, augmented variations of the attack prompt, submits them in parallel, and scores/yields the strongest bypass output.
Launch Sandbox JobIndirect Prompt Injection (RAG/CWE-1158)
Embeds hidden adversarial directives into retrieved external context (HTML/Markdown/Search) to hijack search assistants and RAG agents.
Launch Sandbox JobPDF Document Structure Injection
Exploits document extraction pipelines by embedding instruction overrides inside PDF metadata, annotations, or invisible text layers.
Launch Sandbox JobMulti-Modal Vision Prompt Injection
Injects adversarial typographic instructions into visual document scans, low-contrast image layers, or OCR streams for visual LLMs.
Launch Sandbox JobPliny & BT6 GG Adversarial Methodology
Demonstrating gradient-guided suffix optimization (BT6 GG) combined with Pliny Liberator structural prompt synthesis. Evaluates cross-architecture jailbreak transferability across frontier models.
Rigorous Defensive Evaluation
All benchmarks are evaluated using standard parameters (budget limits, temperature=0.0, HeuristicClassifier taxonomy scoring) to ensure reliable replicability. We strictly avoid model damage and promote constructive disclosure.