laurentkempe.comarticleAgentic Workflows
A personal blog post walks through building a GitHub Agentic Workflows automation (a GitHub Next research demonstrator) that autonomously checks for new Astro releases, researches breaking changes, applies upgrades, and opens a PR — with no human involvement until review. The piece details the tool's security model in depth: agents get read-only repo permissions by default, all writes must go through pre-declared 'safe-outputs' (e.g. create-pull-request with a compile-time allowed-files list), external network access is restricted via an allowlist enforced by an 'Agent Workflow Firewall', and PRs are never auto-merged so a human always reviews the change.
Organisations named here
- Continuous AI
- Don Syme
- GitHub Next
- Idan Gazit
- Peli de Halleux
Technologies named here
- GitHub Agentic Workflows
- Astro
- @astrojs/mdx
- gh aw compile
- @astrojs/sitemap
- GitHub Copilot
- Agent Workflow Firewall
- astro-auto-import
- safe-outputs
- astro-icon
- @astro-community/astro-embed-twitter
- @astro-community/astro-embed-youtube
- @tailwindcss/vite
- tailwindcss
- GitHub Actions
- pnpm / pnpm-lock.yaml
- npm
- pnpm
Techniques named here
- Prompt injection resistance via restricted write channels
- Least-privilege / read-only default permissions
- Network allowlisting
- File-scope restriction for automated changes
- Human-in-the-loop review gate
Related in AI Security
blog.talosintelligence.comtoolIntroducing CAIRN: Frontier tracking for AI-integrated malware (Cisco Talos)
github.comadvisoryFlowise MCP Security Bypass Enables RCE
github.comrepoYesWeHack Claude Kit
jfrog.comarticle