GitHub's security model assumes human users, not prompt-injectable agents. Three design issues, a GraphQL exfiltration vector found by a Claude pentest, and the layered defenses needed to close them.
Airut now supports scheduled tasks -- two modes for running Claude Code on a schedule, enabling automated CI fixes, daily briefings, and pipeline refinement.
Building software, managing accounting, and deploying diagnostics — all by emailing Claude Code through Airut.
How writing about agentic AI security revealed a sandbox escape via GitHub Actions, building the fix, and pentesting it with Claude.
Why agentic AI security can't be solved deterministically, how the industry's current approaches fall short, and two criteria for evaluating practical security.
How we extended our diagnostics tool from a single-pass analysis pipeline into an agentic workflow where customer support iterates with an AI agent over email to resolve issues.
How Varjo's commercial XR technology became the default choice for high-end military simulation — a dual-use story running in the opposite direction.
How emailing tasks to Claude Code replaced the terminal interface — and became an open-source project called Airut.
How LLMs changed software engineering for me in 2025 — building a 100k-line tool, measured impact, and observations on working with AI coding agents.
What we observed from piloting Claude Code on a large C++ codebase — quantitative results, use cases, limitations, and ROI.