← Back to all news
Published on · by Renaud Deraison

Webinar with GitGuardian: Your Coding Agent Has the Keys

An hour with GitGuardian on hardening developer workflows when coding agents need real access: how that access is granted, tracked and contained, and how Bromure keeps real credentials out of the agent's reach.

Thank you to GitGuardian for inviting me to their webinar Your Coding Agent Has the Keys: Hardening Developers' Workflows. The full recording, about an hour, is on YouTube.

What it's about

Coding agents need real access to write, test and ship code, so the secrets problem can't be designed away. What you can decide is how that access is granted, tracked and contained. The session is for AppSec and DevSecOps engineers who want something more concrete than "write an AI policy".

GitGuardian's State of Secrets Sprawl 2026 shows why it matters now:

  • Developers using AI coding assistants leak secrets at roughly twice the baseline rate.
  • Leaked AI service credentials grew 81% year over year, with 1.27 million exposed in 2025.
  • 24,008 unique secrets turned up in public MCP configuration files alone.

Attacks like s1ngularity and the Shai-Hulud variants already use AI command-line tools to steal secrets and spread.

Protecting against a rogue agent starts with limiting the access you give it. It also means closing off the access it can grab on its own: the plaintext credentials sitting on the machine, outside the sandbox. That is what Bromure Agentic Coding is built for. Claude Code, Codex and Grok Build run in isolated Linux VMs, and the real credentials never enter the agent's reach.

An earlier Bromure, the same principles

The demo shows an earlier version of Bromure. The interface has changed a lot since, but the principles haven't: each agent works in a VM of its own, the credentials inside it are stand-ins, and your Mac swaps in the real ones only for the services they belong to. To see Bromure as it is today, watch the episodes or read the manual's chapter on credentials.

GitGuardian also covers Agent Hooks, which scan for secrets in what AI coding tools read and write.