Skill UI

This guide details how to implement robust content safety guardrails for applications powered by Claude. It covers essential techniques such as system prompt hardening, input validation to detect injection patterns and length limits, and output validation to prevent system prompt leakage. It ensures compliance with Anthropic's policies, making the AI integration secure and reliable.

A comprehensive evaluation suite designed to benchmark the performance of Large Language Models (LLM) and Vision-Language Models (VLM) specifically for home security applications. It covers 143 diverse tests across 16 critical domains, including tool use, prompt injection resistance, contextual reasoning, alert triage, and detailed scene analysis, ensuring robustness for smart home AI systems.

A comprehensive offensive checklist for assessing the security and robustness of AI and Large Language Model (LLM) applications. It covers advanced adversarial techniques such as prompt injection, jailbreaking, model extraction, data poisoning, and analyzing system vulnerabilities across components. Essential for red-teaming and security assessment.

A comprehensive expert playbook detailing advanced Large Language Model (LLM) prompt injection techniques. Covers direct methods (instruction override, role play), indirect attacks (RAG poisoning, web data manipulation), and sophisticated tool/function calling abuse for data exfiltration and security bypass. Essential for rigorous AI application security testing.

This skill outlines the implementation of advanced runtime defenses for production Large Language Model (LLM) applications. It covers using multiple industry-leading guardrail systems—including Llama Guard, NeMo Guardrails, and LLM Guard—to detect and block adversarial attacks such as prompt injection, jailbreaks, and toxic content. It provides a comprehensive defense-in-depth strategy for ensuring model safety and output integrity in real-world deployments.

A comprehensive security tool designed to detect and defend against subtle prompt injection attacks hidden within untrusted multi-modal content, such as web pages, PDFs, and images. It scans for obfuscated payloads (e.g., hidden CSS, metadata, zero-width characters) by combining content normalization, heuristic analysis, and dedicated ML detectors, ensuring that LLM agents only consume sanitized, trustworthy input.

PyRIT (Python Risk Identification Tool) is an open-source automation framework designed for advanced, multi-turn adversarial red-teaming of conversational LLMs. It simulates real-world attack scenarios by orchestrating an attacker model and a scorer model in a feedback loop. Features include Crescendo (gradual escalation) and Tree-of-Attacks with Pruning (TAP) techniques to detect complex vulnerabilities like prompt injection and jailbreaks in stateful dialogues.

This skill utilizes NVIDIA's open-source garak framework to conduct comprehensive red-teaming assessments on Large Language Models (LLMs). It tests for critical vulnerabilities such as prompt injection, jailbreaks, data leakage, and toxic content generation by sending thousands of adversarial probes. Ideal for pre-deployment security validation, API guardrail testing, and generating defensible evidence for AI risk assessments.

This skill provides a comprehensive defense-in-depth framework for securing autonomous AI agents. It implements least-privilege tool allowlisting, identity binding, and human-in-the-loop controls to mitigate risks like prompt injection and tool poisoning. Use this when building agents that interact with sensitive systems (e.g., payments, file writes) to ensure governance, auditability, and compliance.