Introducing SafeLLM v1.0: Enterprise AI Security Gateway
Today we're launching SafeLLM, the first open-source security gateway designed specifically for LLM applications. Learn about our defense-in-depth approach.
Firewall for Your LLMs
SafeLLM transforms Apache APISIX into an intelligent AI security gateway. Multi-layer protection against prompt injection, PII leaks, and abuse. Zero Trust for AI · GDPR/SOC2 Ready · ~6-16ms latency
<16ms
Latency
850+ RPS
Throughput
>89%
Accuracy
80%
Cost Savings
Defense-in-Depth
Waterfall Security Pipeline
Each request passes through multiple security layers. Short-circuit principle: dangerous requests are blocked immediately, saving computational resources.
L0: Semantic Cache
If a similar question was already asked, SafeLLM returns cached response from Redis, bypassing the model. Saves up to 80% of API costs.
L1: Keyword Guard
Blazing fast (O(1), FlashText) blocking of known jailbreak patterns and system commands. Fully configurable phrase lists.
L1.5: PII Shield
Dual Mode: Fast regex (1-2ms) or precise AI (GLiNER, 20-25ms). Enterprise: define custom entities (employee IDs, project numbers).
L2: AI Guard
Neural networks (ONNX) detecting sophisticated prompt injection attacks. Classes: safe, jailbreak, indirect_injection.
DLP Output Scanning
Scans model responses. If the model "spits out" confidential data, SafeLLM blocks or anonymizes it.
Shadow Mode
Safe Day-0 deployment: log "would_block" but allow all requests. Tune thresholds before going live.
Air-Gapped & High Availability
Enterprise Ready
Built for the highest infrastructure security requirements
100% Offline / Air-Gapped
Enterprise version works completely without internet access. All AI models (ONNX/GLiNER) loaded locally. Your data never leaves your network.
High Availability (HA)
Redis Sentinel support (cache failover) and Distributed Coalescer (K8s pod coordination) ensures protection continuity even during node failures.
Custom Model Support
Replace standard filters with powerful Guard-class models (Llama, Gemma, Qwen architectures) with GPU acceleration for SOTA detection.
Compliance & Audit
EU AI Act Ready, RODO/GDPR, SOC2/ISO 27001. Non-editable audit logs (Loki/S3). Full data sovereignty in your VPC.
Dashboard & Observability
Visual security testing interface. Prometheus + Grafana integration. Token ROI Dashboard showing cost savings.
Sidecar Pattern
Native integration with Apache APISIX. Docker & Kubernetes ready (Helm Charts). Stateless & horizontally scalable.
Pricing
Open Source vs Enterprise
Start free with OSS, scale with Enterprise
0
Popular
Enterprise
Full Power
$ Custom
contact us- Everything in OSS
- AI Guard (ONNX L2)
- PII GLiNER (25+ entity types)
- Air-Gapped Mode
- Redis Sentinel HA
- Distributed Coalescer
- Dashboard & Audit Logs
- GPU Guard Models
- Priority Support
Latest from the Blog
View all posts »Security insights, tutorials, and updates from the SafeLLM team.
Prompt Injection 101: Understanding the Threat
Prompt injection is one of the most critical security vulnerabilities in LLM applications. Learn how attackers exploit it and how to defend against it.
PII Protection for LLMs: GDPR Compliance Made Simple
How to prevent sensitive data from reaching cloud LLM providers. A practical guide to PII detection and anonymization in AI workflows.
Secure Your AI Infrastructure Today
Join companies protecting their LLM deployments with SafeLLM.
Start with Open Source or schedule an Enterprise deep-dive.