Breaking

Curated Microsoft Security skills for AI agents released on GitHub

July 8, 2026 at 20:25 EDT

AI Agents
Security
Open Source

A "skills" package that embeds Microsoft Security expertise into AI coding agents has been released on GitHub as open source under the MIT license. Covering areas such as Defender, Sentinel, Entra, Purview, Intune and Security Copilot, it packages skills grounded exclusively in public Microsoft Learn documentation, to be loaded into existing agents like GitHub Copilot, Claude Code, Cursor, Codex CLI and Gemini CLI.

Open Source · MIT License

Microsoft Security expertise, packaged into your AI coding agent

A "skills" package of 88 curated security skills — grounded exclusively in public Microsoft Learn docs — plugs into agents like Copilot, Claude Code, Cursor, Codex CLI and Gemini CLI to deliver opinionated, guardrailed guidance instead of generic advice.

curated security skills, grouped by purpose

AI hosts supported (Copilot, Claude, Cursor…)

test prompts · 61 assertions across 10 high-risk skills

Behavioural harness: assertions passed (of 61)

Baseline vs. with skills — two frontier models, no regressions observed.

Claude Opus 4.8

84%

baseline 51/61

100%

+skills 61/61

GPT-5.5

90%

baseline 55/61

100%

+skills 61/61

The 88 skills, grouped by purpose

Threat protection / SecOps

Defender XDR, Sentinel

Identity / governance

Entra ID, Conditional Access, PIM

Compliance / data

Purview DLP, eDiscovery, Insider Risk

AI / agent security

Agent 365, Copilot oversharing, DSPM

Endpoint / device

Intune, BitLocker, PAW

Cloud / platform

Azure Policy, Key Vault, WAF

SecOps acceleration

Security Copilot agents, recovery

Welcomed as

"A strong step toward operationalizing security knowledge inside AI workflows" — embedded decision trees and guardrails could close the gap between generic recommendations and production-grade actions.

Caution

Freshly released — real-world validation remains scarce. Some skills may need tuning, and comprehensive production benchmarks or failure reports have not yet surfaced.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.