Lakera Red

Getting Started with Lakera Red

Lakera Red provides comprehensive AI security assessments to identify vulnerabilities in your GenAI applications. This guide walks you through running your first security scan.

Prerequisites

Before you begin, ensure you have:

  • A Lakera account with Red access enabled
  • Access to your GenAI application’s endpoint or model configuration
  • System prompt and configuration details for your application (optional but recommended)

Access the Red Platform

  1. Navigate to the Lakera Red Platform
  2. Sign in with your Lakera credentials
  3. You’ll see the Red dashboard with your organization’s scans and targets

Core Concepts

Before running a scan, understand these key concepts:

ConceptDescription
TargetA reusable configuration for the system you want to test (model or API endpoint)
ScanA security assessment run against a target
Recon ContextInformation about your app that helps generate targeted attacks
Attack ObjectiveA specific security test (e.g., “system prompt extraction”)
Scan ResultThe outcome of testing one attack objective

Create a Target

Targets are reusable configurations that define how Red connects to your GenAI application.

2

Choose connection type

Select your target type:

  • Model - Direct connection to a model (e.g., GPT-4, Claude)
  • Stateless Endpoint - API endpoint with bearer token authentication
3

Configure connection

Provide the required details:

  • Target name and description
  • Model name or endpoint URL
  • System prompt (if applicable)
  • Authentication credentials
  • Any extra parameters
4

Test connection

Click Test Connection to verify Red can communicate with your target before saving.

Run Your First Scan

1

Start a new scan

Go to Scans → New Scan

2

Choose your target

Select an existing target or create a new one.

3

Provide recon context

Help Red generate targeted attacks by describing your application:

  • App Description - What does your application do?
  • Allowed Actions - What should the AI be able to do?
  • Forbidden Actions - What should the AI never do?
4

Select security test scope

Choose which attack categories to include:

  • Security - Instruction override, prompt extraction, data exfiltration
  • Safety - Harmful content generation, dangerous instructions
  • Responsible - Misinformation, copyright, fraud facilitation

You can also select specific attack objectives within each category.

5

Launch the scan

Click Launch Scan to start the assessment.

Monitor Scan Progress

After launching, you’ll be taken to the progress page where you can:

  • Watch the live feed of attack completions
  • See real-time stats: elapsed time, attacks completed, issues detected
  • Continue working while the scan runs in the background

Scan statuses:

  • preparing → testing → evaluating → completed
  • Scans may also end in failed, timeout, or cancelled

Review Your Results

Once the scan completes, view your results in two ways:

By Risk Category

See results grouped by attack category (security, safety, responsible), with risk scores for each.

By Test

See results grouped by individual attack objective, showing which specific tests found vulnerabilities.

For each result, you can view:

  • The conversation - exact prompts sent and responses received
  • The evaluation - why the attack was considered successful or not
  • The attack success score (0-5, where 3+ indicates a successful attack)

Understanding Risk Scores

Your scan produces a risk score representing the percentage of harmful evaluations:

Risk ScoreSeverity
≤25%Low
26-50%Medium
51-75%High
>75%Critical

Export Results

Export your scan results for reporting or further analysis:

  • JSON - Full results with all metadata, conversations, and evaluations
  • CSV - Flattened format with objective names, scores, and explanations

Next Steps

Need Help?

For questions about Lakera Red or to discuss your assessment needs, contact our team.