Our team will be out of office on Friday, May 1, 2026. We’ll be back and ready to assist you starting Monday, May 4th.

What Does My AI Coding Agent Actually Send Back To The Company That Made It?

Contents

A plain answer to the question almost no founder asks before installing an AI agent: what does it read, what does it send, and where does that data land?


You gave an AI agent access to your files. You typed a prompt, watched it work, and felt that jolt of “this changes everything.” I have felt it too. I feel it most weeks.

Now answer me one question. Where did your files go?

Here is the direct answer: your AI coding agent sends more than you think, usually more than it needs, and in at least one documented case published this month, it sent the entire repository even when it was explicitly told not to read anything. On July 10, 2026, a security researcher publishing under the handle cereblab released a wire level teardown of xAI’s official Grok Build CLI. Using a proxy to capture the tool’s own network traffic, the researcher documented that the CLI transmitted the contents of files it read verbatim and unredacted, including a .env secrets file, and separately uploaded a snapshot of the whole repository, git history included, to a Google Cloud Storage bucket named grok-code-session-traces. That upload happened independently of what the agent actually read.

I want to be precise, because precision matters more than outrage here. The researcher was careful to state what was not proven. There is no proof that xAI trains on that data. Upload is not training. Storage is not training. What was documented is transmission, acceptance, and storage. That is the honest scope, and I am going to hold to it.

I am not telling you to stop using AI agents. I use them every single day. They have made me faster and better at my work, and they have done the same for tens of thousands of entrepreneurs I have taught. This is not a “be afraid” article.

This is a “know your plumbing” article.

The moment you point an agent at a folder, you have created a data pipeline out of your company. That pipeline is real whether or not you thought about it. Almost no founder has ever traced where it goes. Before you install another agent, you need three answers: what does it read, what does it send, and where does that land.

That is the whole thesis. Let me show you why it matters and exactly how to get those three answers.

Key Takeaways

  • Giving an AI agent file access creates a data pipeline out of your business, and you are responsible for knowing where that pipe ends.
  • A July 2026 wire level analysis documented xAI’s Grok Build CLI transmitting unredacted .env contents and uploading entire repositories to a named Google Cloud Storage bucket, though it did not prove xAI trained on that data.
  • IBM’s 2025 Cost of a Data Breach report found organizations with high shadow AI usage averaged $670,000 in higher breach costs, and 97 percent of breached organizations with an AI security incident lacked proper AI access controls.
  • The fix is not abstinence from AI. The fix is a repeatable audit you run in about twenty minutes before you trust a new tool.
  • Every agent you deploy should get scoped credentials, a limited workspace, and an audit trail, the same way you would treat a new contractor.

The Problem

Here is what I have to admit to you, because I would rather be useful than look clever.

For a long stretch of the last two years, I installed AI tools the way most of you do. Fast. A curl command from a homepage, a browser login, a folder pointed at it, and off to the races. I read the marketing page. I did not read the terms. I certainly did not run a packet capture. I am a founder and a teacher, not a security researcher, and I made the same silent assumption you probably made: a reputable company would not move my data anywhere I would not expect.

That assumption is not malice on anyone’s part. It is just an assumption. And assumptions are exactly what get audited last.

The problem is structural, and it is worth naming honestly. Cloud AI agents have to send your context somewhere. That is not a scandal; that is how they work. Anthropic’s Claude Code ships with a one million token context window. Large context agents are designed to read a lot of your material. That is the product, and it is a good product.

So the question was never “does it send anything.” The question is the delta. What is the gap between what the model needed to do the job and what actually left your machine? In the Grok Build teardown, the researcher measured that gap directly. On a 12 GB test repository the agent never read, the model channel moved about 192 KB. The storage channel moved 5.10 GiB. That is roughly a 27,800 times difference between what the model ingested and what left the machine.

That gap is the whole ballgame. It is the difference between a tool doing its job and a tool doing something else on the side.

And here is the part that stings: you cannot see it. There is no light on the dashboard. No prompt. No confirmation dialog. The agent works beautifully, you get your result, and the pipe runs quietly in the background.

Now the reframe, because there is one. This is not a reason to retreat. It is a reason to get good at something almost nobody is good at yet. Founders who can answer the three questions have an advantage in every enterprise conversation, every partnership, every client contract with a data clause. Knowing your plumbing is becoming a competitive credential.

The Evidence

Let me give you the receipts, plainly, with the caveats intact.

The Grok Build CLI teardown, July 2026. The researcher used mitmproxy to capture the tool’s own traffic on their own machine, using a throwaway repository with fake canary secrets. Findings: file contents including a .env were transmitted verbatim and packaged into an archive that uploaded successfully. Separately, with the explicit prompt “Reply with exactly: OK. Do not read or open any files,” the CLI uploaded the entire repository as a git bundle. Cloning that captured bundle recovered a planted file the agent had been told not to open, marker verbatim, plus the full commit history. The result replicated on a second unrelated repository. The researcher also reported that turning off the “Improve the model” setting did not stop the upload. Read section 7 of that gist. The author lists what was not proven, and that intellectual honesty is why I am willing to cite it.

xAI has since responded. As of July 13 and 14, 2026, reporting from The Register and others indicates the codebase upload was disabled server side via a flag, and that xAI has said Grok Build respects zero data retention for customers who enable it, with a command available to delete previously collected data. Elon Musk publicly stated that uploaded user data would be deleted. I am including this because fairness requires it. The behavior was documented; the company responded. Both things are true.

Shadow AI has a measurable price tag. IBM’s 2025 Cost of a Data Breach report, researched by the Ponemon Institute across 600 organizations, found that organizations with high levels of shadow AI averaged $670,000 more in breach costs than those with low or no shadow AI. One in five organizations reported a breach connected to shadow AI. Only 37 percent had policies to manage AI use or detect shadow AI. And 97 percent of breached organizations that had an AI related security incident said they lacked proper AI access controls.

Your team is already feeding these tools. Cyberhaven Labs, which instruments actual enterprise endpoints rather than surveying opinions, reported that 34.8 percent of corporate data employees put into AI tools is sensitive, up from 27.4 percent a year earlier and 10.7 percent two years before that. The most common category of sensitive material going in is source code, at 18.7 percent.

This is not new, and the pattern repeats. In 2023, Samsung engineers pasted proprietary semiconductor source code and a confidential meeting transcript into ChatGPT; within roughly twenty days of allowing the tool, three separate incidents occurred, and Samsung banned external generative AI on company devices. In August 2023, Zoom revised its terms of service twice in one week after users read a clause that appeared to permit training on customer content. This month, Meta launched Muse Image on July 7, 2026, defaulted public Instagram accounts into a feature that let strangers generate AI images from their photos, and pulled it days later after backlash, saying it “missed the mark.”

Notice the through line. In none of these cases did anyone hack anything. The data moved because of a default, a term, or a design choice. And in every case, somebody had already agreed to it.

The Solution

Here is what changed for me.

I stopped treating AI tools as software and started treating them as employees. Not metaphorically. Operationally.

When you hire a contractor, you do not hand them a master key to the building and walk away. You give them a badge that opens two doors, for ninety days, and the badge system logs every swipe. That is not distrust. That is how grown up businesses run. I had somehow decided that logic did not apply to a binary I installed from a curl command, and that was the error.

The framework I now run, and the one I teach inside White Beard Strategies, is deliberately short enough that you will actually use it. Three questions. I call it the Read, Send, Land audit.

Read. What is inside the blast radius? Not what the agent will probably touch. What it can touch. If you launch an agent from your home directory, the answer is your entire life. If you launch it from a scoped project folder with no credentials in it, the answer is much smaller. This one question, answered honestly, eliminates most of the risk before you ever look at the network.

Send. What leaves the machine, and is it more than the job requires? You do not need to be a security engineer to check this. You need to look for the delta the Grok teardown made visible: the gap between what the model consumed and what the tool uploaded. Vendors who publish this clearly are telling you something. Vendors who are vague are also telling you something.

Land. Where does it come to rest, who can read it, for how long, and can you delete it? “We do not train on your data” and “we do not store your data” are entirely different promises. Ask for both in writing. The Grok case is the cleanest possible illustration: the training opt out and the upload behavior were two separate systems, and switching one off did not touch the other.

I am not asking you to become paranoid. I am asking you to spend twenty minutes before you spend two years of trust.

The good news is the industry is building the tools for exactly this, and it is happening fast. Google published Mantis Skills, an open, portable toolkit for building security review harnesses around coding agents. AWS is running workshops on agent identity and access management, and their framing is blunt: long lived agent credentials are a breach waiting to happen. Their recommendation is scoped, short lived credentials through STS, automatic rotation through Secrets Manager, and audit trails through CloudTrail. Microsoft’s production agent playbook assigns each agent a distinct identity and its own workspace.

Read those three together and the consensus is obvious. Agents get identities. Agents get workspaces. Agents get logs. The people building this stuff already know. The gap is that almost none of it has reached the founder running a lean team with no security hire, and that gap is exactly what we work on with our members.

Practical Steps

1. Inventory every agent with file access today. Open a document and list every AI tool on your machines that can read local files: coding CLIs, desktop apps, IDE extensions, MCP servers, browser agents. Most founders find three to five more than they expected. You cannot audit a pipe you do not know exists.

2. Run the Read question on each one. For every tool, write down the directory you actually launch it from and what lives under it. If the answer includes credentials, client data, or your accounting exports, you have found your first fix, and it costs you nothing to make it.

3. Move secrets out of the blast radius. Get .env files, API keys, and credential exports out of any folder an agent can reach. Use a secrets manager or environment injection at runtime. The Grok teardown proved a .env can be transmitted unredacted; assume every agent does the same until proven otherwise.

4. Give every agent its own scoped, short lived credentials. No shared master keys, no permanent tokens. Follow the AWS pattern: scoped permissions, automatic rotation, and a logged trail. If an agent goes sideways, you want to revoke one badge, not rebuild your whole key infrastructure at midnight.

5. Read the data section of the terms, and only that section. Skip the rest. You are looking for three phrases: what is retained, for how long, and whether training is opt in or opt out. Ten minutes. Do it before you install, not after an incident.

6. Ask your vendor the Land question in writing. Send one email: “Where is my content stored, who can access it, what is the retention period, and how do I request deletion?” A good vendor answers in a day. The response time itself is data.

7. Write a one page AI usage policy for your team. Only 37 percent of organizations have one. It does not need to be legal grade. It needs to say which tools are approved, what data never goes into an AI tool, and who to ask. That single page moves you out of the expensive shadow AI column.

Frequently Asked Questions

Is my AI coding agent training on my private code?

Usually not by default on paid business tiers, but training and storage are different questions. A tool can decline to train on your data while still uploading and retaining it. Ask both questions separately, get both answers in writing, and never assume that opting out of training stops the upload.

Does this mean I should stop using AI coding agents?

No. I use them daily and I recommend them constantly. The productivity gain is real and the founders who abandon these tools will lose to the founders who use them carefully. The goal is scoped access and known plumbing, not abstinence. Use the tools with your eyes open.

What is shadow AI and why does it cost so much?

Shadow AI is any AI tool your team uses without approval or oversight. IBM’s 2025 report found organizations with high shadow AI averaged $670,000 more in breach costs. The expense comes from having no visibility: you cannot control access, log usage, or respond quickly to an incident you never knew existed.

How do I check what a tool sends without being a security engineer?

Start with the questions, not the packets. Read the vendor’s data documentation, email them the retention question, and check whether independent teardowns exist. If you want the technical layer, Google’s open Mantis Skills toolkit is built for exactly this kind of agent security review.

My business is not technical. Does any of this apply to me?

Yes, and arguably more. The same pattern shows up in note takers reading your meetings, assistants reading your inbox, and image tools reading your photos, as Meta’s Muse Image rollout demonstrated this month. Read, Send, Land works on every AI tool, not just coding agents.

The Close

Go back to that moment. The one where you pointed an agent at your files and felt the future arrive.

I want you to keep that feeling. I mean that. That feeling is correct. These tools are the biggest leverage advantage independent entrepreneurs have ever been handed, and I have spent years of my life helping people grab it. Nothing in this article walks that back.

But leverage cuts both ways, and the same access that makes an agent brilliant is what makes it a pipe. You built that pipe. You built it with a curl command and a login and a folder, in about ninety seconds, and you agreed to whatever was on the other end without reading it. Not because you were careless. Because nobody told you that was the moment that mattered.

Now you know.

So here is what I want from you, and I want it this week, not someday. Open the document. List the agents. Ask the three questions. Move the secrets. Write the page. It is twenty minutes of work that separates the founders who get to use these tools at scale, in front of real clients, under real contracts, from the founders who are one incident away from losing the right to use them at all.

The winners in this era will not be the people who moved fastest. They will be the people who moved fast with their eyes open. Speed without sight is just falling.

Point the agent wherever you want. Just know where the pipe goes.

If you want the audit worksheet, the vendor question templates, and the one page policy we hand our members, that is exactly what we build together inside White Beard Strategies training. Come get it.


About the author: Jonathan Mast is the founder of White Beard Strategies, where he teaches AI coaching and mentorship to a community of more than 100,000 entrepreneurs. He is the creator of the Perfect Prompt Framework, a widely used method for getting reliable, business grade output from AI tools, and he is a frequent speaker on practical AI adoption for founders and small teams. Jonathan’s work focuses on one idea: entrepreneurs do not need to become engineers to win with AI, they need systems they can actually run. He writes and teaches every week at whitebeardstrategies.com.

About the Author