🛡️

PromptGuard

LLM security assessment and red teaming tool. Evaluates AI applications for prompt injection, data leakage, and policy violations — producing risk scores and structured findings reports for hardening production LLM deployments.

Active Python FastAPI LLM Security Red Teaming

PromptGuard — LLM Security & Red Teaming

Overview

PromptGuard is a security testing tool for LLM-powered applications. As AI becomes embedded in production systems, the attack surface grows beyond traditional software security — prompt injection, data exfiltration via model outputs, and policy bypasses represent a new class of risk that standard security tooling does not cover.

The tool runs a structured suite of probes against a target LLM endpoint, evaluates responses for security indicators, and produces a risk-scored findings report with actionable remediation guidance. Built to be used as part of a security review before deploying AI features to production.

Sample Findings

⚠️Prompt InjectionHigh Risk

⚠️Data Leakage via OutputHigh Risk

⚠️Jailbreak SusceptibilityHigh Risk

⚡System Prompt ExtractionMedium Risk

⚡Role Confusion ExploitMedium Risk

Architecture

Probe EngineStructured test suite of security probes, each targeting a specific vulnerability class. Probes are templated and parameterised — new probe types can be added without modifying core logic.

EvaluatorResponse analysis layer. Each model response is evaluated against indicators for the probe's target vulnerability. Produces a binary finding plus a confidence score.

ScorerComposite risk score (0–100) aggregated from individual finding scores with severity weighting. Score bands: Low / Medium / High / Critical.

ReporterStructured findings report: risk score, per-finding detail, affected probe, severity, and remediation guidance. Output in JSON and human-readable format.

APIFastAPI interface for integration into CI/CD pipelines or manual assessment workflows. POST target endpoint, receive full findings report.

Key Features

Structured probe suite covering injection, leakage, jailbreak, role confusion, and policy bypass
Composite risk scoring — weighted aggregate across all findings with severity bands
Structured findings report — per-finding detail with severity, description, and remediation
CI/CD integration ready — REST API accepts target endpoint and returns findings as JSON
Probe library extensible — add new vulnerability classes without modifying the evaluation core
Historical scan storage — compare findings across versions to track security posture changes

Tech Stack

Python FastAPI LLM APIs asyncio SQLite Pydantic

Demo Preview

▸ PROMPTGUARD — SECURITY SCAN