Skill UI

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 automatically enriches research wiki pages that are left as bare scaffolds after initial ingestion. It reads multiple external sources (like arXiv/alphaxiv) and fills in 10 structured `_TODO._` sections (e.g., Problem/Gap, Key Results, Open Questions) with detailed prose summaries. Use this skill to transform incomplete research scaffolds into a comprehensive, structured knowledge base.

This skill uses Python and boto3 to analyze AWS CloudTrail logs. It establishes statistical baselines of normal API activity and detects various anomalies, such as unusual event sources, first-time API usage, geographic IP deviations, and high-frequency calls. This is critical for identifying credential compromise, privilege escalation, and unauthorized resource access in a security incident.

Automated system designed to monitor paste sites (e.g., Pastebin, GitHub Gists) for leaked credentials, API keys, and sensitive data dumps. It uses advanced scraping, keyword matching, and regex patterns to provide early detection of data breaches and security incidents.

This skill provides a repeatable assessment for RAG systems, testing for critical vulnerabilities in vector stores and embedding models. It specifically addresses embedding inversion, cross-tenant data leakage, and knowledge-base poisoning, aligning with OWASP LLM08:2025. Use when performing authorized security audits of retrieval layers and multi-tenant isolation.

This skill audits Model Context Protocol (MCP) servers to detect critical AI security vulnerabilities. It scans for tool poisoning, tool shadowing, rug pulls, SSRF, and unauthenticated exposure. Use it before integrating new tools or during security reviews to ensure agent reliability and prevent supply chain attacks.

Deploy and operate Falco with the modern eBPF driver in Kubernetes and Docker. Covers driver selection, Helm installation, output channels, and the built-in ruleset that detects container escape, namespace abuse, privileged mounts, and anomalous syscalls. Ideal for enabling runtime security, routing alerts to SIEM or Falcosidekick, and upgrading existing deployments.

A comprehensive tool for identifying poisoned training data, backdoored ML models, and supply-chain compromises. It leverages IBM's Adversarial Robustness Toolbox (ART) for activation clustering and spectral signature analysis, uses Cleanlab for label quality checks, and enforces integrity via weight-hash verification and safetensors. Essential for securing ML pipelines before training, deployment, or during supply chain audits.

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.

This skill provides comprehensive methods and frameworks to detect sophisticated AI attacks, including model extraction, membership inference, and model inversion, which abuse inference APIs. It focuses on monitoring per-principal query volume, input distribution, and confidence exposure. Use it for securing public or partner inference APIs, or for conducting pre-deployment red-teaming to assess the model's extractability and overall security posture against MITRE ATLAS techniques.

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.