Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward

2025-11-25 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Cybersecurity & Data Privacy · Depth: Advanced, extended

Summary

This survey details the rapid evolution of agent skills for large language models (LLMs), which enable dynamic capability extension without retraining by providing composable packages of instructions, code, and resources. The field is organized into four axes: architectural foundations, including the SKILL.md specification and Model Context Protocol (MCP); skill acquisition methods like reinforcement learning with SAGE, autonomous discovery with SEAgent, and compositional synthesis; deployment within the computer-use agent (CUA) stack, highlighting advancements in GUI grounding and benchmarks like OSWorld; and security, revealing that 26.1% of community-contributed skills contain vulnerabilities. The paper proposes a Skill Trust and Lifecycle Governance Framework, a four-tier, gate-based permission model, to address these security concerns and outlines seven open challenges for realizing trustworthy, self-improving skill ecosystems.

Key takeaway

For AI Scientists and CTOs deploying LLM agents, the shift to modular agent skills necessitates a robust security strategy. Given that 26.1% of community skills contain vulnerabilities, you must implement a multi-tiered governance framework, like the proposed Skill Trust and Lifecycle Governance Framework, to manage skill provenance, verification, and runtime permissions. This approach is critical to mitigate prompt injection, data exfiltration, and privilege escalation risks, ensuring the integrity and safety of your agentic systems.

Key insights

Agent skills modularize LLM capabilities, extending functionality dynamically while introducing significant security challenges.

Principles

• Skills provide procedural knowledge, not just tool execution.
• Progressive disclosure minimizes context window consumption.
• Trust tiers should align permissions with skill provenance.

Method

The proposed Skill Trust and Lifecycle Governance Framework uses four verification gates (static analysis, semantic classification, sandboxing, permission manifest validation) to assign skills to graduated trust tiers, with continuous runtime monitoring for lifecycle evolution.

In practice

• Use SKILL.md for structured, dynamically loaded instructions.
• Implement multi-stage verification for community-contributed skills.
• Consider capability-based permissions for agent tools.

Topics

• Agent Skills
• Large Language Models
• Computer-Use Agents
• Model Context Protocol
• Agent Security

Code references

• scienceaix/agentskills

Best for: AI Scientist, Research Scientist, CTO, AI Researcher, AI Engineer, AI Security Engineer

Related on AIssential

• DAY 3 Livestream - 5-Days of AI Agents: Intensive Vibe Coding Course With Google — Progressive disclosure of agent skills prevents context rot by loading minimal metadata and fetching full instructions only when needed.
• SkillWiki: A Living Knowledge Infrastructure for Agent Skills — SkillWiki provides a unified infrastructure for agent skill management, from ingestion to evolution, linking skills to evidence.
• tech-leads-club / agent-skills — Agent Skills provides a secure, validated, and extensible library of capabilities for AI coding agents.
• A Primer on Using Agent Skills — Agent skills provide a modular, dynamic way to equip AI agents with specific capabilities, improving efficiency and reliability.
• Agent Skills Work but the Research Shows Most Teams Are Building Them Wrong — Curated, focused agent skills improve AI performance, but require structured organization and security diligence to scale effectively.
• Security and Privacy of Large Language Models: Threat Taxonomy, Ethical Implications, and… — LLM risks are systemic, requiring a lifecycle-oriented defense-in-depth strategy across technical, human, organizational, and governance layers.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.