Product overview

The LLM safety scanner built for regulated environments

LLMSec runs 30+ adversarial probes against any LLM endpoint and produces a compliance-mapped PDF report. It runs entirely inside your own infrastructure — your VPC, your Kubernetes cluster, your air-gapped server.

Scanner Core

Async probe runner with bounded concurrency. 30+ probes across 5 categories. OpenAI-compatible, HuggingFace local, and Ollama adapters. Tenacity-based retry with exponential backoff.

Compliance Report

20–40 page PDF report generated locally. Every finding mapped to SR 11-7 (Fed/OCC), OWASP LLM Top 10, NIST AI RMF, and EU AI Act. Ed25519-signed audit trail on every report.

Offline License System

Ed25519-signed license keys verified entirely offline. No phone-home required. Trial, Standard, and Enterprise tiers enforced cryptographically. Air-gapped deployments supported.

CLI Interface

Customer-facing primary interface. Rich progress output, JSON and text modes, probe list with filtering, scan history. Security engineers live in CLIs — we built for them.

Probe library

30+ probes across 5 categories. Each probe declares its name, category, severity (CRITICAL / HIGH / MEDIUM / LOW), and framework mappings. Probes are inspired by and in some cases adapted from Garak (MIT license, credited).

Jailbreak

~10 probes

Direct instruction override
Role-play persona bypass
Goal hijacking
Multi-turn escalation

Prompt Injection

~5 probes

Direct injection
Indirect injection via retrieved content
System prompt override

Data Extraction

~5 probes

System prompt leakage
PII extraction
Training data memorization

Toxicity

~5 probes

Hate speech elicitation
Discriminatory output
Harmful content generation

Reliability

~5 probes

Refusal consistency
Multi-turn coherence
Behavioral stability

CLI usage

# Run a full scan

$llmsec scan --target openai:gpt-4o-mini --output json > results.json

# Filter to specific probe categories

$ llmsec scan --target openai:gpt-4o-mini \

--probes jailbreak,prompt_injection,toxicity \

--output json > results.json

# List available probes

$ llmsec probe list

$ llmsec probe list --category jailbreak

# Scan local model via Ollama (free, no API key)

$ llmsec scan --target ollama:llama3.2

Model adapters

LLMSec scans any LLM — not just the ones from specific vendors. The adapter interface is the abstraction layer that makes this possible.

Adapter	Target format	Coverage
OpenAI Compatible	openai:<model>, azure:<model>	OpenAI, Azure OpenAI, Anthropic, any OpenAI-compatible API (~80% of targets)
HuggingFace Local	huggingface:<model-id>	Any HuggingFace Transformers model running locally
Ollama	ollama:<model>	Any model served by Ollama — free, no API key required