RedForge
github.com
- Free plan
- Linux
- API
- Self-hosted
What RedForge is
RedForge is an open-source platform for red teaming and vulnerability scanning of large language models. It lists more than 49 probes, covers the OWASP LLM Top 10, and lets users add YAML probes without Python. Attack modules include PAIR, Crescendo multi-turn escalation, Skeleton Key, many-shot jailbreaking, and GCG adversarial suffixes. Its mutation engine expands two payloads into 37 variants using encoding, Unicode, cognitive reframing, and structural injection strategies. Provider adapters are listed for OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Mistral, Ollama, HuggingFace, and generic REST endpoints. The Ollama adapter can run locally without an API key. Scans produce JSON, JSONL audit, and failures reports, with HTML, Markdown, and SARIF also available. Compliance reporting covers NIST AI RMF, EU AI Act, ISO/IEC 42001, GDPR, and OWASP LLM Top 10, with custom YAML frameworks supported. RedForge can emit SARIF for GitHub Code Scanning and return status 1 on critical or high findings for CI gating. It is self-hosted, available for Linux and through an API, and released under Apache 2.0. Use requires an authorization value indicating owned, authorized, or research use.
Who it suits
RedForge suits teams and researchers conducting authorized LLM security assessments, particularly those who want self-hosted scans or CI integration. Its YAML probes and custom compliance frameworks also suit users who need to extend the test or reporting scope.
Good to know, and what to watch
- More than 49 listed probes
- Covers the OWASP LLM Top 10
- Supports local Ollama use without an API key
- SARIF output supports GitHub Code Scanning
- Custom YAML frameworks are supported
- CLI requires an authorization value
- REST API limits clients to 10 scans per minute
- Available platforms list Linux and self-hosted use
Plans
Every plan and priceRedForge
Open source
The verdict
RedForge brings probe-based testing, multiple model-provider adapters, and compliance reporting into a self-hosted tool. Its authorization requirement and API rate limit are worth accounting for when planning scans.
What the maker publishes
- Attack methods
- Attack modules include PAIR, Crescendo multi-turn escalation, Skeleton Key, many-shot jailbreaking, and GCG adversarial suffixes. source
- Attack modules
- Its attack modules include PAIR, Crescendo multi-turn escalation, Skeleton Key, many-shot jailbreaking, and GCG adversarial suffixes. source
- CI integration
- RedForge can emit SARIF for GitHub Code Scanning and exits with status 1 on critical or high findings for use as a CI gate. source
- Compliance
- Compliance reporting covers NIST AI RMF, EU AI Act, ISO/IEC 42001, GDPR, and OWASP LLM Top 10, and users can add custom YAML frameworks. source
- Deployment
- The Docker instructions say the container runs as non-root user redforge (UID 1001) with a read-only filesystem. source
- Detectors
- The README lists seven detector types, including YARA-style, similarity, refusal, regex, keyword, code, and unsafe-content detectors. source
- Guardrails
- A bidirectional guardrail pipeline scans inputs and outputs for prompt injection, PII, toxicity, and secrets. source
- Integrations
- It can export SARIF for GitHub Code Scanning and documents a GitHub Actions workflow that uploads the SARIF file. source
- License
- The project is released under the Apache 2.0 license, and the README says probe payloads are released under CC0. source
- License and use
- The repository identifies the license as Apache 2.0 and says the tool is for authorized use only. source
- Local use
- The Ollama adapter runs locally without an API key, and the README describes it as zero cost. source
- Model providers
- It lists adapters for OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Mistral, Ollama, HuggingFace, and generic REST endpoints. source
- Mutation engine
- The mutation engine expands two payloads to 37 variants across encoding, Unicode, cognitive reframing, and structural injection strategies. source
- Probe coverage
- It lists 49+ probes and full OWASP LLM Top 10 coverage, with YAML probes that can be added without Python. source
- Providers
- The README lists provider adapters for OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Mistral, Ollama, HuggingFace, and generic REST endpoints. source
- Purpose
- RedForge is an LLM security testing platform for red teaming and vulnerability scanning. source
- Reports
- Scans automatically produce JSON, JSONL audit, and failures reports, with additional HTML, Markdown, SARIF, JSON, audit, and failures formats. source
- Security controls
- Documented controls include masking API keys in logs, setting result files to mode 0600, bearer-token REST API authentication, disabled-by-default CORS, and a rate limit of 10 scan requests per minute per client. source
- Usage requirement
- The CLI requires an authorization value indicating owned, authorized, or research use. source
Checked against the maker's pages on 9 Oct 2026.