Skill UI

Get a deep, multi-round, and adversarial critical review of your research work, ideas, or drafts. This skill simulates the perspective of a senior ML reviewer (e.g., NeurIPS/ICML level) using dedicated external backends (Codex or Manual). It is designed to find logical gaps, missing experiments, and narrative weaknesses, providing brutal honesty to elevate your academic contribution.

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.

This tool audits the integrity of machine learning experiments by checking for common fraud patterns such as fake ground truth, score normalization fraud, and phantom results. It utilizes an external reviewer backend (like Codex or manual review) to perform an adversarial review across all project artifacts, ensuring reported results are trustworthy and methodologically sound.

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.

This skill performs rigorous mathematical proof verification using an adversarial, cross-model review process. It reads a LaTeX proof, systematically identifies logical gaps (e.g., unjustified assertions, quantifier errors) and theoretical flaws (e.g., illegal interchanges, non-uniform convergence). It then iteratively fixes these gaps with full derivations, re-reviews, and generates a comprehensive audit report, ensuring mathematical rigor and completeness. Use this when seeking deep, academic verification of a theory paper.

A comprehensive guide detailing advanced security threats in Machine Learning and AI systems. Covers model supply chain attacks (RCE via pickle, poisoning), adversarial examples (FGSM, PGD, physical-world), data privacy breaches, and LLM-specific vulnerabilities. Essential for building robust, secure AI pipelines.

Open Dynamic Workflows (ODW) is an advanced, open-source engine designed to manage complex coding tasks using multiple AI agents. It operates in three stages: planning the task into parallel subtasks, orchestrating various coding agents (like Codex or Antigravity) to work concurrently, and finally, performing an adversarial verification pass to challenge and validate the combined output before merging. It supports multiple model providers and is ideal for structured, high-reliability software development.

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.

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 is designed for complex, multi-step tasks requiring coordination among multiple AI agents. It implements a robust workflow engine that first generates a detailed plan (task graph, topology) and then orchestrates parallel execution with adversarial verification. Use this when the task spans many files, requires multiple specialized roles, or involves an overall 'workflow' concept, ensuring structured, reliable, and auditable results.

This skill runs comprehensive benchmarks against structural and adversarial corpora. It measures the performance and cost efficiency of a booster model compared to leading LLM baselines (Gemini, Anthropic). Results are persisted in verifiable JSON reports, crucial for auditing claims and verifying booster win rates before publishing.