Skill UI

Handles IMO/Putnam/USAMO/AIME style problems by generating refined proofs, then running a fresh-context adversarial verifier to catch flaws, finally producing honest confidence reports and verified PDFs when possible.

Use Clarvia AEO (Agent Experience Optimization) to score any MCP server, API, or CLI tool for agent-readiness. It evaluates 15,400+ AI tools across four dimensions—API accessibility, data structuring, agent compatibility, and trust signals—ensuring the components are robust enough for advanced AI agent workflows and dynamic selection processes.

A comprehensive tool designed to systematically plan ablation studies for machine learning research papers. It guides the process by having an AI (Codex) simulate a rigorous reviewer, identifying necessary experiments to isolate component contributions and answer critical research questions. It also includes a feasibility review step to manage compute resources and implementation complexity.

A multi-layered defense system designed to detect sophisticated prompt injection attacks targeting LLM-powered applications. It combines regex pattern matching for known signatures, heuristic scoring for structural anomalies, and advanced DeBERTa-based classification. Ideal for building robust input validation layers, securing RAG pipelines, and performing proactive AI security scanning.

This skill detects AI-generated deepfake audio used in voice phishing (vishing) attacks. It functions by extracting critical spectral features (including MFCCs, spectral centroid, and zero-crossing rate) and passing them through advanced machine learning models (like Random Forest and Gradient Boosting). The system supports batch forensic analysis, provides confidence scores, and generates detailed reports to verify the authenticity of recorded voice calls, making it essential for cybersecurity incident response and fraud investigation.

This framework detects NTLM relay attacks across Active Directory environments by correlating Windows Security Event 4624 (LogonType 3). It identifies key indicators such as IP-to-hostname mismatches, Responder/LLMNR poisoning artifacts, and non-enforced SMB/LDAP signing, providing comprehensive threat detection for T1557.001.

This skill is designed for performing local, on-premise evaluations and benchmarking of models hosted on the Hugging Face Hub. It facilitates running specialized evaluation frameworks like `inspect-ai` and `lighteval` against local hardware. Users can select optimal inference backends, including `vllm` for high throughput, or fall back to Hugging Face Transformers or `accelerate`, ensuring comprehensive smoke testing and task selection for LLMs. It focuses purely on the local execution process, separate from remote jobs or publishing workflows.

This tool automates the end-to-end workflow for GDPR Data Subject Access Requests (DSAR). It handles identity verification, discovers PII across structured and unstructured data sources using NER and regex, maps data to Article 15 requirements, and generates auditable compliance reports, ensuring adherence to ICO/EDPB guidelines and scaling compliance efforts.

This framework implements robust input and output validation guardrails for LLM-powered applications. It prevents prompt injection, detects and redacts PII, filters toxic content, and validates model outputs against defined schemas. Use it when deploying AI agents that handle sensitive data or require strict adherence to corporate policies, ensuring a secure and trustworthy AI interaction layer.

Run Hugging Face models directly in JavaScript/TypeScript for text, vision, and speech tasks without backend servers, using Transformers.js in Node.js or modern browsers with WebGPU, quantization, and pipeline abstractions for efficient inference.

A comprehensive skill for managing, provisioning, and destroying GPU instances on vast.ai. It analyzes your deep learning task to determine precise hardware requirements (VRAM, GPU count, storage) and searches for cost-optimized, high-reliability GPU offers. The skill handles the entire lifecycle from initial analysis to experiment monitoring, ensuring you only rent what you need at the best possible value.

Assesses AI/ML systems for prompt injection, jailbreak, model inversion, data poisoning, and agent tool abuse using MITRE ATLAS mapping, signature detection, and scoring metrics before deployment.