An infra repo full of secrets

DevOps, SRE and securityOps team with kubeconfigs, tfvars and vault files in reach

Our infra repo has kubeconfig files, terraform.tfvars and a .vault-token in reach of the agent. Before turning on a memory layer, our security lead wanted proof that none of it would leave a laptop.

Will an AI agent memory tool send my kubeconfig or tfvars files off my laptop?

Run arcroom sent --all to see exactly what ArcRoom sent in the last 7 days: only metadata, never file contents, tool output or diffs. A secret file's path, or a command that names one, leaves as [secret file]. Use arcroom off to turn it off for one repo. Prompt redaction is by pattern, so keep secrets in files or env vars, not inline.

Also asked as: Is it safe to give an agent memory layer access to an infra repo with secrets? · How do I check what data an agent hook sends from my machine? · Can I prove our kubeconfig and vault token never leave the laptop?

The work

Only metadata is sent. File contents, tool output and diffs never are.

ArcRoom keeps
  • Redaction
  • Off switch
The next agent gets

$ arcroom sent --all

Step by step

  1. Only metadata is sent. File contents, tool output and diffs never are.
  2. A secret file's path, or a command that names one, leaves as [secret file].
  3. Prompts are cut at 1,000 characters and secrets in them are redacted by pattern.
  4. arcroom sent shows what left in the last 7 days; arcroom off turns ArcRoom off for one repo.

What the agent sees

$ arcroom sent --all
sent: { tool: 'Read', path: '[secret file]' }
sent: { cmd: 'kubectl --kubeconfig [secret file] get pods' }
0 file contents

Compared with today's tools

No memory layer in the infra repo

Leave infra work out of ArcRoom and keep notes by hand.

Setup
None
Ongoing
re-explain context every session

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init), check with arcroom sent
Ongoing
None

Staying off is the safest choice if even metadata can't leave; ArcRoom lets a team see what is sent and decide per repo.

Where it stops

Redaction is by pattern. A token pasted inline in a prompt or command in an odd format could slip through, so keep secrets in files or env vars.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in devops, sre and security