PromptRAID
github.com
- Free plan
- Windows
- Linux
What PromptRAID is
PromptRAID is a free, open-source command-line framework for testing prompt injection and jailbreak behavior in LLM-based agents. It is designed for self-hosted red-team work on Linux and Windows, and maps implemented mutation categories to MITRE ATLAS technique IDs. Its mutation engine creates varied payloads through encoding changes, jailbreak or role-play framing, delimiter breaks, and indirect instructions delivered through tool or document output. The harness supports Anthropic, Gemini, Groq, Cerebras, and OpenRouter, including OpenAI-compatible endpoints. A judge labels outcomes as success, partial compliance, fail, or inconclusive, while separating neutral quoting from following an injected instruction. Runs and individual payload results are stored in SQLite, with JSON and HTML report export. The CLI includes commands for taxonomy, mutation, judging, test runs, demos, and reports. PromptRAID requires Python 3.9 or later and at least one supported provider API key. Some categories remain unverified or lack mutation generators.
Who it suits
PromptRAID suits security researchers and teams red-teaming LLM agents for prompt injection and jailbreak behavior. It is intended for users comfortable configuring provider credentials and running command-line software.
Good to know, and what to watch
- Free and open source under the MIT License
- Supports Linux and Windows
- Maps implemented categories to MITRE ATLAS IDs
- Exports JSON and HTML reports
- Stores runs in SQLite
- Requires Python 3.9 or later
- Needs at least one provider API key
- Data leakage tests have no mutation generator
- Some categories remain unverified placeholders
The verdict
PromptRAID provides a focused command-line workflow for generating, running, judging, and reporting prompt-injection tests. Its provider coverage and MITRE ATLAS mapping are useful, but users should account for the required credentials and incomplete mutation coverage.
What the maker publishes
- ATLAS coverage
- Implemented mutation categories map to verified MITRE ATLAS technique IDs, while some listed categories remain unverified placeholders or have no generator yet. source
- Installation
- The README recommends pipx for CLI use and documents virtual-environment setup, including a Windows activation command. source
- Installation
- The README recommends installing the CLI with pipx and also documents installation in a Python virtual environment. source
- Intended users
- The maker identifies as an independent offensive security researcher, and the project is designed for prompt injection testing and red teaming of LLM agents. source
- Intended users
- The project describes its intended use as systematic red-team testing and benchmarking of LLM-based agents across providers. source
- Interface
- The project provides a command-line interface with commands for taxonomy, mutation, judging, running tests, demos, and reports. source
- License
- The repository uses the MIT License, which permits use, modification, distribution, and sale subject to its stated conditions. source
- License
- PromptRAID is distributed under the MIT License, which permits use, copying, modification, distribution, sublicensing, and sale subject to its terms. source
- Limitations
- The README says triggered prompt injection and data leakage appear in its taxonomy but do not yet have mutation generators. source
- Maintainer
- The README identifies Shikhali Jamalzade (alisalive) as an independent offensive security researcher and the project author. source
- Maker
- The repository names Shikhali Jamalzade (alisalive) as the author. source
- Model providers
- The harness supports Anthropic, Gemini, Groq, Cerebras, and OpenRouter, including OpenAI-compatible endpoints. source
- Payloads
- Its mutation engine generates prompt injection variants using encoding tricks, role-play framing, delimiter breaking, and indirect injection through tool or document output. source
- Payload testing
- Its mutation engine generates prompt injection variants using encoding tricks, jailbreak framing, delimiter breaking, and indirect injection through tool or document output. source
- Provider credentials
- Running tests requires at least one provider API key, with environment variables documented for Anthropic, Gemini, Groq, OpenRouter, and Cerebras. source
- Providers
- The harness supports Anthropic, Gemini, Groq, Cerebras, and OpenRouter, including OpenAI-compatible endpoints. source
- Purpose
- PromptRAID is an automated prompt injection red-teaming framework for LLM-based agents, mapped to MITRE ATLAS. source
- Purpose
- PromptRAID is an automated prompt injection red-teaming framework for LLM-based agents, mapped to MITRE ATLAS. source
- Reports and storage
- PromptRAID exports JSON and HTML reports and stores runs and per-payload results in SQLite. source
- Reports and storage
- PromptRAID exports JSON and HTML reports and stores runs and per-payload results in SQLite. source
- Requirements
- The package requires Python 3.9 or later and asks users to configure at least one supported provider API key. source
- Security classification
- The project maps prompt injection tests to MITRE ATLAS and describes them as red-team tests for LLM agents. source
- Taxonomy
- Mutation categories map to MITRE ATLAS technique IDs, which the project says it checks against the live ATLAS matrix. source
- Unverified categories
- Three additional categories are described as unverified placeholders and are excluded from the listed ATLAS mapping. source
Checked against the maker's pages on 10 Oct 2026.