← Back to all products

AI Safety & Guardrails Kit

$39

Input/output filtering, toxicity detection, PII redaction, hallucination detection, and content policy enforcement scripts.

36 files
PythonYAMLMarkdownJSONLLM

Product Preview

Try the interactive reader and demo tools below, or get the full product with all content unlocked.

Interactive Reader (Free Preview) ⚙ Try Demo Tools Download Free Sample

File Structure 36 files

ai-safety-guardrails/ ├── LICENSE ├── README.md ├── free-sample.zip ├── guide/ │ └── 01-safety-guide.md ├── guides/ │ └── safety-guide.md ├── index.html ├── interactive.html ├── policies/ │ ├── default_policy.yaml │ └── strict_policy.yaml ├── requirements.txt ├── src/ │ ├── __init__.py │ ├── __pycache__/ │ │ ├── __init__.cpython-312.pyc │ │ ├── groundedness_checker.cpython-312.pyc │ │ ├── injection_detector.cpython-312.pyc │ │ ├── input_filter.cpython-312.pyc │ │ ├── output_filter.cpython-312.pyc │ │ ├── pii_detector.cpython-312.pyc │ │ ├── policy_engine.cpython-312.pyc │ │ └── toxicity_detector.cpython-312.pyc │ ├── groundedness_checker.py │ ├── injection_detector.py │ ├── input_filter.py │ ├── output_filter.py │ ├── pii_detector.py │ ├── policy_engine.py │ └── toxicity_detector.py └── tests/ ├── __pycache__/ │ ├── test_injection_detector.cpython-312-pytest-9.0.2.pyc │ ├── test_injection_detector.cpython-312.pyc │ ├── test_pii_detector.cpython-312-pytest-9.0.2.pyc │ ├── test_pii_detector.cpython-312.pyc │ ├── test_policy_engine.cpython-312-pytest-9.0.2.pyc │ └── test_policy_engine.cpython-312.pyc ├── test_injection_detector.py ├── test_pii_detector.py └── test_policy_engine.py

Documentation Preview README excerpt

AI Safety Guardrails

Input/output filtering, PII detection, prompt injection defense, toxicity

screening, and configurable policy enforcement for LLM applications.

Runs entirely on Python stdlib — no external dependencies required.

Why This Exists

Every LLM application needs safety guardrails, but most teams either:

  • Ship without them ("we'll add safety later")
  • Build ad-hoc regex checks that miss obvious attacks
  • Pay for a hosted safety API that adds latency and vendor lock-in

This toolkit gives you a complete safety stack that runs locally with zero

latency overhead, zero external dependencies, and full auditability. Use it

as your primary safety layer for internal tools, or as a fast pre-filter

before a heavier ML-based classifier for high-stakes applications.

Features

  • PII Detection & Redaction — Regex-based detection of emails, phone numbers

(US/international), credit cards (with Luhn validation), SSN-format numbers,

IP addresses, passport-style IDs, and custom patterns. Context-aware confidence

scoring reduces false positives.

  • Prompt Injection Defense — 4-layer detection: known attack patterns,

structural analysis, encoding/obfuscation detection, and statistical anomaly

scoring. Catches direct overrides, DAN jailbreaks, delimiter injection,

conversation injection, base64 payloads, and unicode homoglyphs.

  • Toxicity Screening — Keyword and heuristic-based detection of threats,

self-harm content, harassment, hate speech, and profanity (configurable).

  • Groundedness Checking — Heuristic analysis of whether LLM output is

supported by provided source material. Catches hallucinated entities, numbers,

and unsupported claims.

  • Policy Engine — Configurable rule-based system that combines all detectors

into unified allow/redact/flag/block decisions. Three built-in presets

(default, strict, permissive) plus custom policy support via JSON/YAML.

  • Input & Output Filters — Unified pre/post-processing wrappers for

the full safety pipeline.

Quick Start

5-Minute Integration


from src.input_filter import InputFilter
from src.output_filter import OutputFilter

input_guard = InputFilter()
output_guard = OutputFilter()

def safe_chat(user_message: str) -> str:
    # Check user input
    input_result = input_guard.check(user_message)
    if not input_result.allowed:
        return input_result.rejection_message

    # Call your LLM with filtered text (PII already redacted)
    response = your_llm(input_result.filtered_text)

    # Check model output

*... continues with setup instructions, usage examples, and more.*

Code Sample .py preview

src/groundedness_checker.py """ Groundedness Checker — AI Safety Guardrails ============================================= Checks whether an LLM's output is grounded in provided source material. Hallucination — where the model generates factually incorrect or unsupported claims — is one of the hardest safety problems in LLM applications. This module provides heuristic checks that catch common hallucination patterns without requiring a second LLM call. Detection strategies (stdlib-only): 1. Named entity coverage: Do entities in the response appear in the source? 2. Numerical consistency: Do numbers in the response match the source? 3. Claim density: High density of specific claims without source backing suggests confabulation. 4. Hedging analysis: Lack of appropriate hedging on uncertain claims. For higher accuracy, combine with an LLM-as-judge groundedness check. Runs on Python stdlib. """ from __future__ import annotations import logging import re from dataclasses import dataclass, field from typing import Any, Dict, List, Optional, Set logger = logging.getLogger(__name__) @dataclass class GroundednessResult: """Result of groundedness analysis.""" response: str source: str is_grounded: bool # True if response appears grounded overall_score: float # 0.0 (hallucinated) to 1.0 (fully grounded) entity_coverage: float # What fraction of response entities appear in source number_consistency: float # How well numbers match between source and response claim_density_score: float # Penalizes high claim density without source backing unsupported_entities: List[str] unsupported_numbers: List[str] details: Dict[str, Any] = field(default_factory=dict) class GroundednessChecker: """Checks whether a response is grounded in provided source material. # ... 271 more lines ...