AI Security Guardrail Gateway Service · Hacking Project

Provide a centralized, policy-driven security gateway that enforces input and output guardrails for all LLM traffic in an organization.

**The problem**
Organizations are embedding LLMs into products without a consistent security layer, exposing them to prompt injection, jailbreaks, sensitive-data leakage, toxic output, and unbounded cost/abuse. Each application team reinvents ad-hoc filters, producing inconsistent, bypassable protection. There is no centralized, policy-driven control plane sitting between applications and models to enforce input/output safety, data protection, and usage limits.

**What you'll build**
Build an AI guardrail gateway that proxies all LLM traffic and enforces layered defenses: prompt-injection and jailbreak detection, PII/secret redaction, output moderation, schema/grounding validation, and per-tenant rate and cost limits. Policies are declarative and centrally managed, with full request/response logging for audit and detection. The gateway aligns to the OWASP LLM Top 10 and MITRE ATLAS, giving security teams one enforcement point and observability layer for all AI usage.

**Objective** — Provide a centralized, policy-driven security gateway that enforces input and output guardrails for all LLM traffic in an organization.

**Expected outcome** — A deployable guardrail proxy service with a policy engine, prompt-injection/PII filtering, and an audit log dashboard, benchmarked against a red-team prompt suite.

**Difficulty** — Advanced — requires building a low-latency proxy, layered detection models, policy engine, and rigorous adversarial testing.

**Primary category** — AI/ML Security; Secondary = AppSec, API Security, Data Loss Prevention

**Security domains** — Primary = AI Security; Secondary = AppSec, DevSecOps, Blue Team

**Technical focus** — Prompt-injection defense, jailbreak detection, PII/secret redaction, output moderation, grounding/schema validation, rate limiting, LLM cost governance, policy engine, observability, adversarial testing

**Core technologies** — Languages: Python, Go/TypeScript; Frameworks/Libraries: FastAPI/Envoy, LangChain/LlamaIndex, guardrail libraries (NeMo Guardrails, Guardrails AI, Rebuff-style detectors), Presidio for PII; Security Tools: prompt-injection classifiers, DLP engines, WAF concepts; Cloud/Infra: Kubernetes, API gateway, Redis, Docker; AI/Data: embedding-based similarity for injection detection, small classifier models; Databases: PostgreSQL, vector store for known-attack signatures

**Where it's valued** — Any enterprise deploying LLM products, AI startups, SaaS, fintech, healthcare AI

**Careers it builds toward** — AI Security Engineer, LLM Security Architect, ML Platform Security Engineer, AppSec Engineer (AI focus), Red/Blue AI specialist

**Framework alignment**
• MITRE ATLAS: AML.T0051 LLM Prompt Injection, AML.T0054 LLM Jailbreak, AML.T0057 LLM Data Leakage, AML.T0048 External Harms — the gateway is a mitigation layer for these techniques
• NIST CSF 2.0: Protect (primary), Detect, Respond
• OWASP: LLM Top 10 (LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure, LLM05 Improper Output Handling, LLM06 Excessive Agency, LLM10 Unbounded Consumption)
• CIS Controls v8: 8 (Audit Log Management), 13 (Network Monitoring & Defense), 16 (Application Software Security)

**Innovation** — Enterprise Product — LLM guardrail gateways are becoming standard enterprise infrastructure, and a policy-driven, multi-tenant implementation has clear commercial and enterprise-readiness value. Scores: Innovation 8/10, Startup 8/10, Enterprise-Readiness 8/10, Patent = Possible.