Skill UI

Performs a comprehensive external audit of a deployed repository's state, going beyond simple code review. It checks for critical production gaps, including RLS misconfigurations, webhook idempotency issues, secret exposure, deployment health, and prompt injection vulnerabilities. Ideal for pre-launch gates and ensuring code stability before going live.

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.

An automated framework for continuous, repeatable red-teaming of LLM applications within CI/CD pipelines. It integrates Promptfoo and DeepTeam to test against recognized vulnerabilities like OWASP LLM Top 10, OWASP Agentic, and MITRE ATLAS (e.g., Prompt Injection, Jailbreaks). This ensures that security vulnerabilities do not regress after code or model changes, providing a critical merge-blocking security gate for production LLM apps.

Implements comprehensive runtime safety defenses for Large Language Model (LLM), RAG, and agent applications. This skill integrates industry-leading guardrail systems—including Llama Guard 3, NeMo Guardrails, and LLM Guard—to inspect and constrain all input and output data. It is essential for production environments requiring protection against adversarial attacks such as jailbreaks, prompt injection (OWASP LLM01), toxic content, and sensitive data leakage.

A comprehensive security skill designed to detect and mitigate indirect prompt injection attacks. It scans untrusted external content—including web pages, PDFs, and images—for hidden, obfuscated instructions before an LLM agent processes them. The pipeline combines content extraction (HTML, OCR), normalization, and dedicated ML models (e.g., LLM Guard, Prompt Guard 2) to ensure data safety.

NVIDIA garak is an open-source LLM vulnerability scanner designed for comprehensive red-teaming and security assessment. It systematically probes large language models (via OpenAI-compatible APIs, Hugging Face, etc.) using thousands of adversarial prompts. It covers critical vulnerabilities such as prompt injection, jailbreaking, data leakage, and toxicity. Use this tool to establish a repeatable security baseline before or after deployment, validate the effectiveness of safety guardrails, or generate quantitative evidence for AI risk assessments.

Implements a comprehensive defense-in-depth framework for autonomous AI agents that interact with external tools. It mitigates risks like prompt injection or tool misuse by enforcing strict tool allowlisting, binding tools to least-privilege identities, requiring human-in-the-loop approval for high-impact actions, and utilizing policy enforcement (like NeMo Guardrails) to bound the blast radius and ensure compliance.

This skill provides comprehensive techniques for security-testing Retrieval-Augmented Generation (RAG) pipelines. It probes for indirect prompt injection, which occurs when malicious instructions are embedded in retrieved documents, and tests for embedding-space manipulation. Using advanced tools like NVIDIA garak, Promptfoo, and PyRIT against vector stores (FAISS, Chroma, etc.), users can validate retrieval guardrails and secure CI/CD gates against adversarial inputs and data leakage.

This skill allows users to extract structured data (text, tables, attributes) from rendered web pages. It enhances traditional web scraping by ensuring all extracted strings pass mandatory AIDefence PII checks and prompt-injection gates before being stored or returned to the model. It supports building reusable extraction templates and running one-shot queries.

This tool performs comprehensive safety scanning on untrusted inputs, such as user submissions, API payloads, or webhook data. It detects sophisticated threats like prompt injection, jailbreak attempts, and adversarial attacks using advanced AI defense mechanisms, ensuring data integrity and security before model processing or execution.

This example demonstrates unsafe or malicious review data inputs, specifically designed for calibration and security testing. It simulates prompt injection attacks and must only be used in controlled environments to validate secure data processing and system robustness.

A specialized downstream skill for CTF-sandbox workflows. Used to analyze complex security challenges driven by AI agent interactions, prompt injection, or infrastructure vulnerabilities. Focuses on tracing control chains involving untrusted input, tool side effects, deployment drift, container runtime issues, and CI/CD supply chain provenance. Ideal for rigorous vulnerability assessment in modern, distributed systems.