> For the complete documentation index, see [llms.txt](https://docs.datumo.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.datumo.com/documentation/documentation-en/redteaming/core-concepts-and-structure.md).

# Core concepts and structure

## Datumo Safety

**Identify safety risks in AI systems through automated red teaming**

The Auto Red Teaming page targets large language models (LLMs) **An automated red-teaming workflow for evaluating safety risks**is. By applying various red-teaming strategies to the seeds of the benchmark dataset, attack prompts are generated, and the target model's **vulnerable risks, which risk taxonomy, and which strategy groups trigger jailbreaks**can be identified using consistent criteria and quantitative metrics.

This section guides you through the full process of using Auto Red Teaming to create red-teaming evaluation tasks, run evaluations, and analyze the results.

### 1. Key Terms

<table><thead><tr><th width="177.125">Term</th><th>Description</th></tr></thead><tbody><tr><td>Benchmark Dataset</td><td>A seed library for attack simulation organized by risk taxonomy.</td></tr><tr><td>Risk Taxonomy</td><td>A harmfulness evaluation classification system defined to verify the safety of AI models from multiple angles.</td></tr><tr><td>Auto Red Teaming</td><td>A process that automatically applies various attack strategies based on the seeds of the selected benchmark dataset and repeats them to explore the model's defense limits and vulnerable points.</td></tr><tr><td>Attack Strategy</td><td>A red-teaming strategy applied to generate attack prompts. It includes 16 specialized strategies derived through Sellectstar's research.</td></tr><tr><td>Attack Set</td><td>The attack items actually used in this evaluation. You can select the entire dataset or randomly sample the needed number of items.</td></tr><tr><td>Target Model</td><td>The target LLM model whose vulnerabilities you want to identify.</td></tr><tr><td>Jailbreak</td><td>Refers to cases where the target model generates harmful or inappropriate responses due to an attack.</td></tr></tbody></table>

### 2. Where can it be used? <a href="#undefined" id="undefined"></a>

* Safety validation before releasing an LLM model
* Risk assessment of AI systems in operation

***

### 3. How does AI Red Teaming work? <a href="#ai-red-teaming" id="ai-red-teaming"></a>

Create evaluation task → Configure attack set → Run automated attack simulation → Analyze results in the dashboard

***

### 4. Next Steps

We recommend using Auto Red Teaming in the following order.

#### 1. [**Select Benchmark Dataset**](https://docs.eval.datumo.com/ko/tutorials/safety/benchmark/)

Check the seeds and risk taxonomy configuration to use for attack simulation.

#### 2. [**Create and run evaluation task**](https://docs.eval.datumo.com/ko/tutorials/safety/star-teaming/run-attack-set)

Select the target model to analyze for vulnerabilities, configure the attack set, and then run automated red teaming.

#### 3. [**Result analysis**](https://docs.eval.datumo.com/ko/tutorials/safety/star-teaming/redteaming-results)

Analyze the model's safety vulnerabilities using metrics such as ASR and score in the dashboard.