NXT1 Daily Tech Briefing — May 31, 2026

AI security research / agentic system threat modeling for security engineers

This Systematization of Knowledge paper formally taxonomizes the attack surface of agentic AI systems, moving beyond individual vulnerability reports to a unified threat model. Agentic systems introduce attack vectors across three layers: data supply chain (context injection, memory poisoning), tool supply chain (malicious tool registration, invocation hijacking), and autonomy exploitation (goal misgeneralization, cascading failures across agent networks).

The paper introduces a formal attack/defense model for tool-augmented LLM agents, distinguishing between attacks on the agent's perception layer (what it sees), reasoning layer (how it plans), and action layer (what it executes). It provides the first unified framework for mapping existing defenses (prompt guards, sandboxing, memory auditing) to specific threat categories — enabling structured threat modeling for production agentic deployments.

This taxonomy is likely to become the reference model for agentic AI threat modeling in security engineering; watch for MITRE to incorporate agentic-specific attack patterns into ATT&CK and for red-teaming firms to release agentic-specific playbooks derived from this framework.

MCP infrastructure engineering / enterprise AI agent governance tooling

Version 1.24.0 (May 2026) shipped Federation 2.0 and the first phase of Skills — making this the most-starred open-source MCP control plane with OAuth authentication, dynamic tool discovery, and Keycloak/Entra ID integration. It addresses the core enterprise MCP gap: transforming scattered MCP server chaos into governed, auditable tool access for both autonomous agents and AI coding assistants.

Federation 2.0 enables hierarchical MCP gateway topologies — a parent gateway can route tool discovery requests across child gateways in separate security domains, maintaining a unified tool catalog while respecting org-level auth boundaries. This is the missing piece for enterprises with multiple business units that want central MCP governance without a monolithic deployment.

Federation patterns will become a default requirement in enterprise MCP RFPs by Q4 2026; watch for the Skills phase 1 implementation to mature into a full agent capability registry — the equivalent of a package manager for agentic tool permissions.

Applied AI research / agent engineering for senior AI practitioners

This actively maintained repository curates 2026 AI agent research papers across agent engineering, memory architectures, multi-agent evaluation, workflow design, and autonomous systems — serving as a living literature review for practitioners who need to track the research frontier without reading every arXiv preprint. Coverage spans finance, healthcare, manufacturing, and government deployments.

The curation emphasizes papers that explain HOW agentic systems work at the architectural level — memory curation models, orchestration trade-offs, production evaluation benchmarks — rather than capability announcements. The repository structure by topic (memory, evaluation, workflows) makes it practical as a reference for applied AI engineering decisions.

As the research-to-production gap in agentic AI narrows, curated practitioner repositories like this will become the primary knowledge management tool for AI engineering teams; watch for organizations to fork and maintain internal versions tied to their specific domain agent stacks.

MCP protocol adoption metrics / AI infrastructure benchmarking for engineering teams

41% of surveyed software organizations report limited or broad MCP production deployment (Stacklok 2026 software report), with 97M+ monthly SDK downloads across Python and TypeScript. More than 9,400 public MCP servers are registered, and adoption spans ChatGPT, Cursor, Gemini, VS Code, and Microsoft Copilot — the fastest protocol adoption curve in enterprise AI infrastructure history.

The report breaks down adoption by deployment pattern — local dev (dominant), self-hosted enterprise (growing), and managed cloud MCP (emerging). The gap between local and enterprise production adoption reveals that authentication, auditability, and multi-tenant isolation are the blocking factors rather than protocol understanding or tooling availability.

Production MCP adoption is likely to double by end of 2026 as the enterprise gateway ecosystem matures; the metric to watch is managed cloud MCP hosting by hyperscalers — once AWS or Azure offers native MCP Gateway as a managed service, enterprise adoption will accelerate significantly.