llama-guard

The course provides guidance on best practices for prompting and building applications with the powerful open commercial license models of Llama 2.

LLM Security Project with Llama Guard

Open-source red-team harness for LLMs. Jailbreaks, prompt injection, agentic tool-abuse and image attacks, scored with validated judges (Llama Guard 3, StrongREJECT).

A local, zero-dependency fuzzer for open-weight LLM guardrails (Llama Guard, Prompt Guard). Measures per-family bypass rate and over-blocking - and verifies each bypass still works on the model behind the guard.

Talent is everywhere; exposure isn't. Marg is a voice-first WhatsApp guide for the people around a rural Indian child (parents, teachers, older students), in Hindi, Hinglish, English and Braj. Six Meta models, every message safety-checked in and out. Meta hackathon 2026, Indie track.

Dynamic hybrid RAG that measures and picks the best strategy per document: local embeddings, ChromaDB/Qdrant, cross-encoder reranking, GraphRAG (NetworkX/Neo4j), agentic RAG, Presidio PII masking, Llama Guard, automated evaluation.

Production-grade AI Governance Gateway & Security Ingress proxy leveraging Llama-Guard-3-8B and serverless hardware execution loops (Baseten + Fireworks AI) to enforce real-time perimeter protection, token threat auditing, and circuit-breaker token failover redundancy.

Sub-millisecond native C++20 & CUDA security guardrail engine for production LLM serving (vLLM / TensorRT-LLM). Detects prompt injections, jailbreaks, PII leakage, and GPU token toxicity in under 0.5ms.

MuZeBOS — Meta AI Zero Trust Business Open Server: AI workers draft, you approve. Self-hosted, open-source business OS; agents draft Meta inbox replies (WhatsApp, Instagram, Messenger), posts, ads and GST invoices, and nothing ships until a human approves the exact payload. Signed webhooks. Next.js + SQLite, free local Llama. Preview.

Brazilian LLM guardrails layer, detects and masks BR PII (CPF, CNPJ, RG, CNH, título de eleitor, PIS) with checksum validation, classifies PT-BR prompt injection via Llama Guard 3, and maps every guardrail to its LGPD article. Python library + Docker REST API sidecar.

Evaluating PAIR, GCG, and Prompt-RS jailbreak attacks against LLaMA-3.1, LLaMA-4, and Qwen3-32B. Two-stage defense pipeline (prompt sanitization + LlamaGuard) achieving 95% Defense Block Rate.

Defense-in-depth input safety for LLMs — perplexity gate + FAISS + ModernBERT + LoRA + Llama Guard 3, behind a deterministic policy gate. 99.88% accuracy, 99.47% jailbreak recall, calibrated confidence, ONNX-optimized. Live demo on HF Spaces.

Enterprise RAG chatbot with namespace-scoped RBAC, hybrid BM25 + vector retrieval, JWT auth, PII redaction, and Llama Guard safety — RAGAS faithfulness 0.91 across 77 automated tests.

SentinelShield: Advanced AI content moderation combining Llama Prompt Guard 2, rule-based filtering, and real-time analysis. Protect your applications from harmful content, prompt injection attacks, and inappropriate material with sub-second response times.

Sub-millisecond C++20 system prompt firewall & token sanitizer for LLM inference engines (vLLM / TensorRT-LLM). Prevents indirect prompt injections, jailbreaks, PII leakage, and malformed inputs at the native execution layer with sub-0.5ms latency overhead.

Create useful automations just by describing what you want in plain English — no scripts, complicated setup, or guesswork.

Security gateway for RAG/LLM apps in banking: Presidio PII pseudonymization (BR), prompt-injection and Llama Guard guardrails, ACL-aware retrieval, LiteLLM, Langfuse, and CI red teaming with promptfoo + garak.

A tabletop robot that finds what you've lost and points at it: fifteen languages, pictures stay home, every model Meta's (Meta Global AI Developer Hackathon, startup track)

Example components for LLM application safety: input checks, Llama Guard integration, output validation and audit logging. Python, with a TypeScript rate-limit example.