Investigating how quantization, pruning, and knowledge distillation affect the trustworthiness and safety of compressed LLMs.

The pioneering work on targeted backdoor attacks that demonstrated how poisoned training data can create invisible trojans in deep learning models.

A guide to the latest AI governance and security scanning developments including Snyk's agent-scan for MCP and skill security, NIST's SP 800-53 Control Overlays for AI Systems, and the Cloud Security Alliance's 243-control AI Controls Matrix.

Providing certified risk bounds for retrieval-augmented generation systems to ensure reliable and safe augmented outputs.

The groundbreaking research showing that large language models memorize and can be made to regurgitate their training data, with profound privacy implications.

A guide to AI security datasets including SafetyPrompts, Do-Not-Answer, JailBreakV-28K, and leaked system prompt collections used for evaluating LLM safety, benchmarking red teaming tools, and understanding real-world AI attack surfaces.

A comprehensive guide to LLM and GenAI red teaming tools including garak, promptfoo, PyRIT, PurpleLlama, and LLMFuzzer for testing language model security against prompt injection, jailbreaks, and adversarial attacks.

An exploration of Microsoft's open-source RAMPART and Clarity tools for securing AI agents during development, plus GitHub Security Lab's Taskflow Agent for AI-powered vulnerability discovery and triage.

Why imitating proprietary language models may not be the shortcut it seems, and what this means for AI safety and capability.

A guide to Google's Agent-to-Agent (A2A) protocol security implications and Hadrian's 2026 Offensive Security Benchmark Report analyzing the explosion of AI-powered offensive security tools and their real-world impact.

An in-depth exploration of AI governance frameworks including NIST AI RMF, ISO 42001, Google SAIF, and OWASP standards for managing AI risk in enterprise environments.

How DataSentinel uses game-theoretic approaches to detect prompt injection attacks in LLM systems, earning IEEE S&P 2025 Distinguished Paper recognition.

An exploration of security vulnerabilities in agentic AI systems and the Model Context Protocol (MCP), including tool poisoning attacks, injection experiments, and comprehensive scanning platforms.

A focused guide on Model Context Protocol security covering MCP scanning tools, secure gateways, context protection, audit extensions, and comprehensive security checklists for MCP deployments.

How COSMIC identifies generalized refusal directions in LLM activation spaces, enabling more precise control over model safety behavior.

How RedCodeAgent automates the discovery of vulnerabilities in code-execution AI agents through intelligent adversarial testing.

An in-depth look at the AI cybersecurity platform race between OpenAI's Daybreak (GPT-5.5-Cyber) and Anthropic's Project Glasswing (Claude Mythos), examining how frontier AI models are reshaping vulnerability discovery and patching.

How concept-specific refusal vectors enable precise steering of language model behavior, offering fine-grained control over what models will and won't generate.

How temporal consistency between video frames can be exploited to detect adversarial attacks against video understanding systems.

The foundational study on adversarial example transferability that made practical black-box attacks against real-world ML systems possible.

How TextGuard provides mathematically provable defenses against backdoor attacks in NLP text classification systems.

How data-free backdoor attacks can compromise machine learning models without requiring access to training data, challenging existing security assumptions.

A comprehensive survey of data poisoning and backdoor attack vectors in machine learning, along with the state-of-the-art defenses.

A guide to AI supply chain security covering AI Bill of Materials (AIBOM), dataset fingerprinting, AI agent discovery, and tools for securing the end-to-end AI development and deployment pipeline.