- Home
- LLM Security Tools
- Inferwall
Inferwall
What it is
Inferwall is a self-hosted AI firewall for checking inputs and outputs in large-language-model applications. It scans for prompt injection, jailbreaks, data leakage, unsafe content, and related threats using heuristic, machine-learning, semantic, and configurable LLM-judge layers. Deployment options include a Python SDK, standalone HTTP server, or transparent reverse proxy. First-party wrappers are listed for OpenAI, Anthropic, LangChain, and FastAPI; additional integrations include Ollama, Hugging Face, LlamaIndex, Pydantic, vLLM, and Mistral. Teams can set allow, flag, or block thresholds for each route and add, override, and version signatures under ~/.inferwall/. The Lite profile is about 8 MB and needs no GPU. Standard adds about 730 MB of models, while Full adds the LLM-judge layer and requires LLM credentials. The site says its Rust heuristic engine runs in under 0.3 ms at p99. The engine uses Apache 2.0, community signatures CC BY-SA 4.0, and signatures map to MITRE ATLAS techniques.
For and against
For
- Self-hosted deployment options
- Multiple detection layers
- Route-level allow, flag, or block controls
- Wrappers for major LLM frameworks
- Lite profile needs no GPU
Against
- No detection system is perfect
- Full profile requires LLM credentials
- Standard profile adds about 730 MB of models
- Users manage deployment themselves
The verdict
Inferwall provides configurable, self-hosted screening for LLM application traffic, with several integration and deployment paths. Its own documentation positions it as one layer in a defense-in-depth approach rather than a complete guarantee.
Who it suits
Inferwall suits teams operating their own LLM infrastructure and wanting configurable input and output screening. Its deployment choices fit SDK, server, and proxy-based architectures.
On the record
- Additional ecosystem support
- The site lists Ollama, Hugging Face, LlamaIndex, Pydantic, vLLM, and Mistral among supported stack integrations.inferwall.com · read 5 Oct 2026
- Caveat
- The site says no detection system is perfect and describes Inferwall as one layer in a defense-in-depth strategy.inferwall.com · read 7 Oct 2026
- Deployment
- It can run as a Python SDK, standalone HTTP server, or reverse proxy.inferwall.com · read 7 Oct 2026
- Detection layers
- It combines heuristic, ML classifier, semantic, and configurable LLM-judge detection layers.inferwall.com · read 7 Oct 2026
- Install profiles
- The site describes Lite as about 8 MB, Standard as including about 730 MB of models, and Full as including those models plus LLM credentials.inferwall.com · read 5 Oct 2026
- Integrations
- The site lists first-party wrappers for OpenAI, Anthropic, LangChain, and FastAPI.inferwall.com · read 7 Oct 2026
- License
- The engine is Apache 2.0 and community signatures are CC BY-SA 4.0.inferwall.com · read 7 Oct 2026
- Licensing
- The engine is licensed under Apache 2.0 and community signatures under CC BY-SA 4.0.inferwall.com · read 5 Oct 2026
- Limitations
- The site states that no detection system is perfect and presents Inferwall as one layer of a defense in depth strategy.inferwall.com · read 5 Oct 2026
- Performance
- The site says its Rust heuristic engine runs in under 0.3 ms at p99 and the Lite profile requires no GPU.inferwall.com · read 7 Oct 2026
- Policy controls
- Users can set allow, flag, or block thresholds per route and extend or override signatures under ~/.inferwall/.inferwall.com · read 5 Oct 2026
- Profiles
- The listed Lite profile is about 8 MB; Standard adds about 730 MB of models; Full adds the LLM-judge layer and requires LLM credentials.inferwall.com · read 7 Oct 2026
- Purpose
- Inferwall scans LLM app inputs and outputs for prompt injection, jailbreaks, data leakage, content safety issues, and agentic threats.inferwall.com · read 7 Oct 2026
- Rules
- Users can override shipped signatures, add their own, and version policies under ~/.inferwall/.inferwall.com · read 7 Oct 2026
- Security mapping
- The site says its prompt injection, jailbreak, prompt extraction, and data leakage signatures map to MITRE ATLAS techniques AML.T0051, AML.T0054, AML.T0056, and AML.T0057.inferwall.com · read 5 Oct 2026
