SLM Security Audit
Guardrail system that scans inputs and outputs for PII leaks, system command injections, and safety violations before model execution.
01 / Overview & Capabilities
Guardrail system that scans inputs and outputs for PII leaks, system command injections, and safety violations before model execution.
Key Features
- Offline regex and semantic PII filters (SSN, credit cards, emails)
- System command injection and prompt jailbreak scanners
- Output evaluator to block harmful, invalid, or off-topic outputs
- Extremely fast CPU footprint — checks query in under 5ms
02 / Standalone OS Harness Installers, CLI & Python SDK
In SLMAgents 2.0, SLM Security Audit executes inside the 10-Plane Pure-CPU Agent Harness (slm_harness). Download a self-extracting Individual Agent + Harness OS Installer for macOS, Windows, or Linux (bundling slm_security + slm_harness + the ./slm CLI), install the full 26-agent desktop suite, or install via Python pip:
1. Individual SLM Security Audit + 10-Plane Harness Installers (All OSes)
slm_security + slm_harness + ./slm CLIslm_security + slm_harnessslm_security + slm_harness + ./slm CLI2. Full 26-Agent Studio & Desktop OS Installers
3. 10-Plane Harness CLI (./slm) & Python pip Installation
# Option A: Install via PyPI or editable source with the 10-Plane Harness $ pip install slm-security $ pip install -e ./slm_harness -e ./slm_security # Option B: Run directly through the Unified 10-Plane Pure-CPU Harness CLI (./slm) $ ./slm ask "Audit @auth_service.py for SQL injection, hardcoded secrets, and CWE vulnerabilities" $ ./slm chat # Interactive Rich TUI with 18 slash commands (/trace, /models, /arena, /undo) $ ./slm serve # 100% Offline Web AI Studio + OpenAI /v1 API at http://127.0.0.1:7860
03 / 10-Plane Pure-CPU Harness Integration (slm_harness)
Rather than executing as an isolated prompt wrapper, SLM Security Audit is natively orchestrated by the SLMAgents 2.0 10-Plane Pure-CPU Harness (via slm_harness.runtime.legacy_bridge and CascadeRouter). Every call benefits from hybrid llama.cpp GGUF (Q4_K_M) + Quantized INT4/INT8 CPU execution, watermarked LRU memory governance, grammar-constrained decoding, and deterministic verification at $0 API cost:
| Harness Plane | Subsystem | How It Powers SLM Security Audit on Pure CPU |
|---|---|---|
| Plane 6 | CascadeRouter (<35ms) |
Stage 2 (OWASP / CVE / SAST Code Audit Signature) & Stage 3 (Embedding Margin) |
| Plane 2 & 5 | MemoryArena & Hybrid CPU Runtime |
Specialist Role: coding (Qwen2.5-Coder-1.5B INT4 / Qwen3-4B GGUF Q4_K_M). Managed inside Watermark LRU MemoryArena (lite 2.6 GB / balanced 4.7 GB / power 8.2 GB) with zero-copy session sharing and automatic OOM tier step-down (8B -> 4B -> 3B -> 0.8B). |
| Plane 7 & 8 | Deterministic Verifier & Governance | Plane 7 Deterministic AST Vulnerability Pattern Scanner + Plane 2 GBNF Severity/Remediation JSON Schema |
| Plane 1, 3 & 4 | Kernel, Artifact Scratchpad & 4-Tier Memory | SQLite WAL CheckpointStore (/fork, /rewind, /undo), content-addressed artifact://<sha12> handles to prevent prompt bloat, and hybrid SQLite FTS5 BM25 + HNSW vector memory. |
04 / Source Checkout (Git Sparse-Checkout)
Clone this agent alongside the 10-Plane Pure-CPU Harness (slm_harness) from the monorepo using Git sparse-checkout — no need to download all 26 agents:
Option 1 — Sparse Checkout with 10-Plane Harness (Recommended)
# 1. Create and enter a new directory $ mkdir slm_security_workspace && cd slm_security_workspace # 2. Initialise empty git repo and add remote $ git init $ git remote add origin https://github.com/t00114218-stack/SLMAgents.git # 3. Enable sparse-checkout for both slm_harness and slm_security $ git sparse-checkout init --cone $ git sparse-checkout set slm_harness slm_security # 4. Pull the 10-Plane Harness and agent source $ git pull origin main
Option 2 — Full Repository Clone
$ git clone https://github.com/t00114218-stack/SLMAgents.git $ cd SLMAgents $ ./slm doctor
Note: After checkout, install the harness and agent in editable mode with pip install -e ./slm_harness -e ./slm_security or download the pre-packaged OS Harness Installer above.
05 / Configuration & API Reference
Constructor Parameters
Instantiate SLMSecurityAudit with performance options:
| Parameter | Type / Default | Description |
|---|---|---|
| model_path | str | None | Explicit path to GGUF (Q4_K_M) or quantized ONNX model weights. If omitted, managed automatically by the 10-Plane Harness MemoryArena. |
| cache_dir | str | None | Directory to store model weights offline. Defaults to ~/.cache/slm-security/. Also settable via SLM_SECURITY_AUDIT_CACHE_DIR. |
| n_threads | int | 4 | CPU thread count for 10-Plane Harness inference (auto-tuned to physical P-cores). Optimize for CPU core count. Also settable via SLM_SECURITY_AUDIT_N_THREADS. |
Methods
| Method Signature | Return Type | Description |
|---|---|---|
sanitize(input_text, system_prompt=None, user_input=None) | dict | Redacts identifiers and returns safety classification mappings. |
Method Parameters (Execution Customization)
All main execution methods accept optional system routing parameters:
| Parameter | Type / Default | Description |
|---|---|---|
| system_prompt | str | None | Optional custom system prompt instruction to override the default system template response parameters. |
| user_input | str | None | Optional additional user-supplied target text variables or contextual keys. |
Quick Start
from slm_security import SLMSecurityAudit
auditor = SLMSecurityAudit()
safe_query = auditor.sanitize(
"SSN is 000-11-2222",
system_prompt="Strict PII redaction rules",
user_input="Skip custom IP addresses"
)
print(safe_query)
Environment Variables
Configure agent parameters globally using environment values:
| Environment Variable | Default | Purpose |
|---|---|---|
| SLM_SECURITY_AUDIT_N_THREADS | 4 | Sets CPU inference execution threads. |
| SLM_SECURITY_AUDIT_CACHE_DIR | ~/.cache/slm-security/ | Default directory to store local GGUF (Q4_K_M) and quantized INT4/INT8 model weights (shared with ~/.slm_harness/models). |
CPU Performance Tuning
To run the SLMSecurityAudit engine efficiently on CPU under 1.5 GB memory footprint:
- Match Threads to Core Count: Set
n_threadsorSLM_SECURITY_AUDIT_N_THREADSto match the physical CPU core count. - Sequential Processing: Avoid concurrent processing when batch files are large.
- Garbage Collection: Clear variables and run
gc.collect()to release model RAM blocks after execution.
Verified Input & Output Logs
Diagnostic execution console response running locally on CPU:
→ INPUT:
"SSN is 000-11-2222"
← OUTPUT:
{
'safe': True,
'sanitized_text': 'SSN is [REDACTED_SSN]'
}