Skill UI

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.

This skill provides comprehensive tools to detect data poisoning and model backdooring attacks throughout the machine learning lifecycle. It checks data provenance, detects label-flip anomalies (using Cleanlab), and analyzes model activations and spectra (using ART) to identify malicious triggers or compromised weights, ensuring ML supply chain integrity.

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.

This skill provides methods and frameworks to detect sophisticated AI attacks, including model stealing, membership inference, and model inversion. It focuses on monitoring query patterns, analyzing confidence exposure, and implementing defensive controls on inference APIs to protect intellectual property and data privacy.

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.

This skill structures a highly controlled, bounded development life cycle for AI agents. It mandates a structured three-phase workflow: Planning (Spec), Implementation (Build), and Verification (Review). It is essential for building complex features from scratch, especially when dealing with ambiguous requirements, high risk, or needing explicit human approval gates to ensure the resulting code meets all defined specifications.

This comprehensive skill provides an index of official blueprints and implementation guides for building full-stack, AI-powered applications using Weaviate. It covers advanced patterns such as Retrieval-Augmented Generation (RAG), agentic workflows, data exploration, and multimodal document search, guiding users through best practices for setup, clients, and frontends.

This skill provides utilities to manage and drain the deliberation queue (brain_frames) of the ACC system. It allows users to list waiting frames, retrieve necessary context, perform deliberation on open tasks, and submit continuations using `acc_act(runtime="continue")`. It is essential for checkpointed work saved by headless runs, ensuring continuous task resolution and calibration against predicted outcomes in complex AI workflows.

This skill provides a hybrid, persistent, and searchable knowledge memory system specifically designed for enhancing AI agents' capabilities. It allows developers to record, retrieve, and search key knowledge, architectural decisions, and operational memories automatically. Use it to build agents with long-term context and reliable knowledge management, ensuring findings are synced back into project documentation.

The Anti-Deception Harness detects manipulative user prompts, such as manufactured urgency, pressure to validate, or undue reliance on authority. It calls an external tool to provide a structured scaffold detailing integrity procedures and counter-evidence, ensuring responses are honest and robust, preventing compromised or sycophantic answers. Ideal for high-stakes decision support.