Hey readers! This week GitHub quietly put a vulnerability scanner one slash command away from every Copilot user, including the free tier. It is not the full CodeQL machine, and knowing exactly where it stops is what makes it useful. Let's dig in.

GitHub Copilot App Now Scans In-Flight Code for Vulnerabilities Before Commit - On July 14, 2026, GitHub shipped the /security-review slash command to the Copilot desktop app, giving every subscriber (Free tier included) AI-driven pre-commit scanning without a terminal or experimental opt-in. - TechTimes

The pitch is shift-left: the command reads your current uncommitted diff and returns prioritized findings scored by severity and confidence, focused on five OWASP-aligned classes (injection, XSS, insecure data handling, path traversal, and weak crypto). Because you are still in the flow, you can apply a suggested fix and re-verify in the same interface while you still have the context in your head.

What matters most here is knowing the boundaries. This is LLM inference on your diff, not rule-based static analysis, and GitHub is explicit that it does not replace the rest of the stack:

Treat it as an early smell test, not a gate. It catches obvious issues cheaply and early; CodeQL, Dependabot, and secret scanning still do the heavy, repository-wide work.

GitHub Code Quality is now generally available - As of today, Code Quality is GA on GitHub Enterprise Cloud and Team, pairing CodeQL's deterministic analysis with AI-assisted detection and Copilot Autofix to flag maintainability and reliability issues before merge. - GitHub Changelog

GitHub says its own teams resolve 67.3% of Code Quality findings before merging PRs, which is the kind of number worth watching against your own workflows. Pricing is $10 per active committer per month plus metered AI usage and CodeQL compute on Actions, and billing begins automatically today (July 20, 2026), so check where it is enabled before it starts charging.

GitHub Copilot CLI Closes Agentic Loophole: Plan Mode Can No Longer Edit Your Files - The July 14 CLI update (v1.0.71 series) hard-blocks workspace-mutating tool calls at the tool-call layer during plan mode, so they are rejected before execution no matter what the model asks. - TechTimes

One caveat to note: MCP and external tools still run as separate processes, so the plan-only guarantee is conditional if you have MCP servers connected.

The Copilot news lands the same week as a sharp reminder of why agent permissions deserve scrutiny.

xAI Open-Sources Grok Build Coding Agent After Cloud Upload Exposes SSH Keys, Repos - Researchers reported that Grok Build (CLI v0.2.93) quietly uploaded whole repositories, including SSH keys and password databases in tracked .env files, to an xAI-controlled Google Cloud bucket. - DevOps.com

SpaceXAI open-sourced Grok Build under Apache 2.0 on July 15, turned data retention off by default, and said it is deleting previously retained code, though independent confirmation is still pending. If you ran Grok Build before July 13, the practical advice is to rotate potentially exposed credentials and audit network traffic. As one researcher put it, making source available after the fact is not a substitute for provable security controls.

That is the issue. If you try /security-review this week, remember what it does not cover and keep the rest of your stack running. See you next time.

Made with ❤️ by Data Drift Press