Not all ads are created equal. And the old, low-performing ones? Those have got to go.

Hightouch can help you refresh your fatiguing ads in just one click. They show you insights on your ad performance so that you can spot the low performers, swap ’em out, and focus your energy on scaling up your high-performing ads.

You can do it all in the Hightouch Ad Studio. Ad Studio’s got the context of your ad performance built right into it to help you:

Turn your tired old ads into potentially high-performing, fresh creative with just a click.

Refresh your ad creative today

Hey innovator,

Picture this:

You open a coding assistant on your laptop.

You ask it to help you fix a bug.

It reads a few files. Solves the problem. Returns an answer.

What you don't see is a second channel — running in parallel, silently — packaging your entire project. Every file. Every folder. Your full commit history going back years. Your passwords. Your API keys. Your database credentials.

All of it, bundled and shipped to a server you didn't know about.

In one documented test, the tool sent 5.1 gigabytes of code in 73 chunks while the task itself needed 192 kilobytes.

That's a ratio of 27,800 to one.

This is what a security researcher named cereblab found when they intercepted the network traffic of xAI's Grok Build CLI — Elon Musk's AI coding tool — and published everything on GitHub yesterday.

It landed on the front page of Hacker News within hours. The internet has not been quiet since.

⚡ What actually happened

The researcher routed Grok Build CLI version 0.2.93 through a network interception tool that captures every packet leaving the machine. What they found was two simultaneous data channels running during every session.

The first was expected: files the agent reads to answer your question, sent to xAI's servers for processing. Any coding assistant works this way. You give it context, it generates a response.

The second was not expected at all.

A separate background channel was quietly bundling the entire repository — every tracked file, full Git history — and uploading it to a Google Cloud Storage bucket called grok-code-session-traces. Not just the files the agent opened. Not just the files relevant to your task. Everything. Including files the agent was explicitly told not to read.

The researcher ran a specific test to confirm this. They told Grok, in the session prompt: "Reply OK, do not read any files." The agent complied. It replied OK. It read nothing.

The background channel uploaded the entire repository anyway.

Then the researcher cloned the captured bundle and recovered a file that had never been touched during the session — a canary file planted specifically to detect this kind of behavior. It came back verbatim. With its full Git history.

The tool was not sending what it needed. It was sending everything you had.

🧠 The detail that made developers stop cold

Grok Build ships with a toggle labeled "Improve the model." Every developer who cares about privacy turns that off. It's the standard signal that says: don't use my data.

Turning it off did nothing.

The researcher disabled it, re-ran the session, and captured the server's response: trace_upload_enabled: true. The upload kept running. All 73 chunks. All returning HTTP 200.

The toggle governs whether xAI can use your data for training future models. It does not govern whether your code leaves your machine in the first place. Those are two different switches. Only one of them is visible to you.

xAI had marketed Grok Build as "local-first" — a term that implies your codebase stays on your machine. The wire-level capture contradicts that description directly.

Your inbox is flooded with newsletters. Your feed is chaos. Somewhere in that noise are the insights that could transform your work—but who has time to find them?

The Deep View solves this. We read everything, analyze what matters, and deliver only the intelligence you need. No duplicate stories, no filler content, no wasted time. Just the essential AI developments that impact your industry, explained clearly and concisely.

Replace hours of scattered reading with five focused minutes. While others scramble to keep up, you'll stay ahead of developments that matter. 600,000+ professionals at top companies have already made this switch.

Join them today, for free.

⚠️ Why this matters even if you've never heard of Grok Build

Most people reading this don't use Grok Build specifically. But this story is not really about one product.

It's about what your software is doing without telling you — and whether you have any practical way to find out.

The credentials sitting in your .env file — database passwords, API keys, payment processor tokens — appeared verbatim and unredacted in the captured traffic. Not as hashed values. Not as references. The actual passwords. Shipped to a storage bucket with no documented retention policy, no stated access controls, and no security advisory from xAI even after the story broke.

As of this morning, xAI has issued no formal statement. No explanation of what the bucket is for. No confirmation of whether repositories already collected will be deleted. The changelog for the version released after the story broke — version 0.2.98 — contains no mention of the upload behavior at all.

The fix that did appear came silently: a server-side flag was flipped remotely, stopping new uploads. One researcher, re-testing the same tool 24 hours after publication, observed zero storage uploads. But the fix has been confirmed on one machine and one account. It is not known whether it is global, permanent, or applied to all subscription tiers.

More importantly: fixing the behavior going forward says nothing about what is already in the bucket.

🔄 The bigger pattern — and why it keeps happening

Two weeks ago, a developer reverse-engineered Claude Code and found hidden Unicode markers tracking which users were routing through Chinese proxies. Anthropic removed them within 48 hours, with no formal written statement for days.

Now a developer has intercepted Grok Build's network traffic and found it uploading entire codebases to an undisclosed storage bucket. xAI removed the behavior via a silent server-side flag, with no statement at all.

The pattern is consistent: a researcher intercepts something that users were not told about, the company removes it quietly, and the explanation comes later — if at all.

AI coding tools are embedding themselves deeper into the most sensitive work that people do. Your source code is your business. In many cases it is your company's most valuable asset — the thing your product is built from, the thing your competitors would most like to have.

The people building these tools understand that. The tools they're shipping are learning that too.

🛠 What to do right now — whether you use Grok or not

If you used Grok Build CLI on any private codebase in the past month, treat every secret in that repository as potentially exposed. Change the API keys. Rotate the database passwords. Revoke and reissue the tokens. Doing nothing because "probably nothing happened" is not a sufficient response when an undisclosed copy of your credentials may be sitting in a storage bucket somewhere.

For any AI coding tool you use — not just Grok, not just this week — run a simple check: what does this tool actually send over the network? Tools like mitmproxy make this visible in minutes. The researcher who found this spent a few hours intercepting traffic. That test is now a standard part of evaluating any tool with access to your code.

If you manage a team, ask the question your security policy should already require: what AI tools are connected to our codebases, and what have we verified about what they send?

Final thought

xAI marketed Grok Build as local-first. A researcher spent an afternoon proving it was something else. The company has said nothing.

That's not a product failure. That's a trust failure. And trust, in software that touches your most sensitive work, is the only thing that matters.

The AI coding tool market is one of the fastest-growing categories in technology right now. Every major lab has one. They all want access to your code so they can help you write better code.

That access is a reasonable trade — if you know what you're agreeing to. This week proved that knowing what you're agreeing to requires more than reading the settings page.

PS: The most damning moment in the researcher's analysis wasn't the 5.1 gigabytes. It was the test where they told Grok not to read any files, and then cloned the uploaded bundle and found the file Grok was told not to open — intact, with its full commit history. The tool sent what you told it not to touch. No explanation for that has been offered.

The Supercharged is aiming to be the world's #1 AI business magazine and is on a mission to empower 1,000,000 entrepreneurs worldwide by 2026, guiding them through the transition into the AI-driven creative age. We're dedicated to breaking down complex technologies, sharing actionable insights, and fostering a community that thrives on innovation, to become the ultimate resource for businesses navigating the AI revolution.

The Supercharged is the #1 AI Newsletter for Entrepreneurs, with 25,000 + readers working at the world’s leading startups and enterprises. The Supercharged is free for the readers. Main ads are typically sold out 2 weeks in advance. You can book future ad spots here.

I'm sending this email because you registered for one of our workshops or our affiliates brought you. You can unsubscribe at the bottom of each email at any time.