Advanced Security Research & Insights
Independent security research on AI agent frameworks, OAuth supply-chain vulnerabilities, and defensive strategies for enterprise AI programs and critical infrastructure.
Latest research
Curated research briefs, tooling, and frameworks for AI and application security teams.
Mapping the LLM Threat Surface for Enterprise Security
A systematic framework for assessing and defending against LLM-specific threat vectors across enterprise environments.
OAuth Supply-Chain Security: From the Salesloft/Drift Breach to Zero Trust
Applied research on OAuth supply-chain vulnerabilities, analyzing the UNC6395 campaign that compromised 700+ Salesforce environments via stolen tokens from Salesloft/Drift. Includes GWAPT-aligned penetration testing methodology.
Autonomous Defense: AI Agent Playbooks for Incident Response
Research on how AI agents can execute defensive playbooks, automate ticket resolution, and augment human responders in security operations.
Research by topic
Explore curated research tracks spanning secure AI development, red team methodology, and governance-ready tooling.
LLM Threats
10 curated resources
Model Context Protocol
5 curated resources
Supply Chain
9 curated resources
Automation & Tools
9 curated resources
Latest insights
Strategic viewpoints for CISOs, AI security leads, and product teams.
Your AI Coding Assistant Reviews Code. It Doesn't Review the Sentence That Came With the Tool
97.1% of real MCP tool descriptions in one academic study contained a defect. None of Claude Code, Codex, or Cursor treat that text as an input surface by default.
Your AI Coding Assistant Will Repeat a Secret Back to You. Here's the Rule That Stops It
Claude Code, Codex, and Cursor all review the code you write. None of them stop the assistant from echoing a credential already sitting in your working set back into chat.
A CLAUDE.md File Will Not Stop an Agent From Pushing a Bad Commit
Client-side rules files are advisory. API-driven agents bypass local git hooks entirely, and a 40-PR overnight credential leak shows exactly how that gap gets exploited.