wehub-resource-sync
0d3cb498a3
4.1 KiB
sidebar_position, sidebar_label, title, description
| sidebar_position | sidebar_label | title | description |
|---|---|---|---|
| 10 | Target Discovery | Target Discovery Agent | Red team AI systems through automated target discovery to detect vulnerabilities, extract system information, and improve attack efficacy in production deployments |
Target Discovery
Promptfoo's Target Discovery Agent automatically extracts useful information about generative AI systems that you want to red team. This information is used to craft adversarial inputs that are unique to the target system, improving attack efficacy and response evaluation quality.
Usage
promptfoo redteam discover
When to use
- CLI: Enhancing the
redteam.purposefield of yourpromptfooconfig.yaml - Self-Hosted: Redteam Target Creation / Usage Details / Application Purpose
- Cloud: Redteam Target Creation / Usage Details / Application Details
In Self-Hosted and Cloud, we find that mapping the answers to the given form fields works best:
| Answer | Self-Hosted | Cloud |
|---|---|---|
| 1. The target believes its purpose is: | Main Purpose | Main Purpose |
| 2. The target believes its limitations to be: | Limitations / Core App Details: Is there anything specific the attacker should know about this system or its rules? | Access and permissions: What systems, data, or resources does your application have access to? |
| 3. The target divulged access to these tools: | Access & Permissions: What systems, data, or resources does your application have access to? | Access and permissions: What systems, data, or resources should your application NOT have access to? |
| 4. The target believes the user of the application is: | User Context: red team user / Who is the red team user? | Access and permissions: What types of users interact with your application? |
How it works
The Target Discovery Agent works iteratively, sending probing questions to the target AI system and evaluating responses until satisfied with the gathered information. This process creates a structured profile for targeted red team attacks.
The agent discovers four key areas:
- Purpose: The system's primary function and intended use cases
- Limitations: Operational constraints, restrictions, and safety guardrails
- Tools: Available external functions, APIs, and their interfaces
- User Context: How the system perceives and categorizes users
The responses are synthesized into a comprehensive profile to inform attack strategies. For privacy, target responses are not stored except in error cases where they may appear in Promptfoo Cloud's error logs for debugging purposes.
Troubleshooting
If promptfoo redteam discover fails with Remote server returned HTTP 400: Unknown task: target-purpose-discovery (or a similar 400/404 error), your client is most likely out of date. Update with npm install -g promptfoo@latest and rerun.
Other common errors:
- HTTP 401 / 403 — Run
promptfoo auth loginand confirm your account has access to target discovery. - HTTP 429 — You are being rate limited; wait a moment and try again.
- HTTP 5xx — The remote generation service is temporarily unavailable. Retry in a few minutes.