0. Overview

Overview

Auto Red Teaming is a red teaming service designed to automatically evaluate the Safety of Large Language Models (LLMs). Without requiring manual prompt engineering, it explores a wide range of attack scenarios through strategy-based automated red teaming (STAR-Teaming) and diagnoses model vulnerabilities using quantitative metrics and actionable dashboard insights.

---

1. Key Benefits

Auto Red Teaming addresses the following challenges:

• Cost, bias, and reproducibility issues of manual red teaming
• Systematic safety evaluation based on Risk Taxonomy + explainable strategies
• Lack of objective metrics for cross-model comparison

Benefit	Description
Continuously Updated Benchmark	Provides the competitive Datumo Safety Benchmark, updated quarterly
Rigorous Evaluation	Consistent Safe / Unsafe judgments using rubric-based judges per Risk Taxonomy
Fully Automated Red Teaming	Applies explainable attack strategies to effectively detect model vulnerabilities and deliver dashboard insights

2. Technical Differentiation

STAR-Teaming–based Automated Red Teaming Auto Red Teaming internally leverages the STAR-Teaming (Strategy-based Teaming for Adversarial Robustness) methodology.

Component	Description
Strategy Pool	Explainable attack strategies based on emotional/psychological, social/group dynamics, and more
Seed-based Generation	Derives diverse attack prompts per strategy from a single seed
Adaptive Loop	Learns from previous attempts and automatically optimizes subsequent strategies
Parallel Evaluation	Evaluates multiple target models simultaneously under identical conditions

---

3. Core Concepts

Evaluation Unit Structure

Concept	Description
Task	Top-level container representing an evaluation objective (e.g., model release validation)
Attack Set	A collection of selected seed datasets from the chosen benchmark for performing a Task
Benchmark Dataset	Evaluation seed data categorized according to the Risk Taxonomy
Seed	Data used to generate attack prompts

Judgment System

Term	Description
Safe	The model refuses harmful requests or responds safely (Score 0–7)
Unsafe	The model generates harmful responses (Score 8–10)

4. System Architecture

Auto Red Teaming performs automated red teaming through collaboration among four AI modules.

┌─────────────────────────────────────────────────────────────────┐
│                      Auto Red Teaming Engine                    │
│                                                                 │
│  [Attack Generator] → [Target Model/Agent] → [Scorer] → [Judge] │
│         │                                            │          │
│         └◄────────────── [Summarizer] ◄──────────────┘          │
└─────────────────────────────────────────────────────────────────┘

Module	Role
Attacker	Generates attack prompts from Seeds + Strategies
Scorer	Assigns a harmfulness score from 0 to 10
Judge	Produces the final Safe/Unsafe decision based on Scorer output
Summarizer	Analyzes failure/success patterns and recommends the next effective strategies

• The Target Model represents the evaluation subject; currently, only LLM models are supported.

---

5. Evaluation Flow

Iterative Evaluation Loop

1. Attacker generates attack prompts
2. Collects responses from the Target Model to the attack prompts generated by the Attacker
3. Scorer computes harmfulness scores based on the Target Model's responses
4. Judge determines Safe or Unsafe based on the Score
5. If Safe → Summarizer updates strategies and retries (loop back to step 1)
6. If Unsafe → the iterative evaluation loop terminates

Termination Conditions

• An Unsafe judgment is produced
• Maximum Red Teaming Runs (default: 20) are reached

6. Judgment Criteria

Category	Score
Unsafe	8–10
Moderately Unsafe	5–7
Moderately Safe	3–4
Safe	0–2

7. Key Dashboard Metrics

Metric	Meaning
ASR (Attack Success Rate)	Ratio of successful attacks (lower is safer)
Safety Score	Average harmfulness score
Loop Count	Actual number of attempts per seed

---