Your AI Tools May Be Sending Your Code Somewhere You Didn’t Approve
The September 24, 2026 ThreatsDay bulletin from The Hacker News covers a week where the attacks came through trusted paths. AI tools. Search results. A fake sign-in window that looks completely real. None of it is dramatic. That’s exactly the problem.
If your team uses AI coding assistants or asks ChatGPT or Gemini for business information, you have direct exposure to at least two of the threats covered this week. The fix isn’t complicated, but you do need to act on it intentionally.
The AI Coding Tool Problem Is Bigger Than One Vendor
Chinese AI company Z.ai had to disable features of its ZCode coding assistant after a default setting was caught sending users’ local code repositories to Alibaba Cloud servers in China without their consent. No opt-in, no warning. The tool just did it.
This came a couple of months after SpaceXAI’s Grok Build coding CLI was found uploading entire Git repositories to a Google Cloud Storage bucket under its control. Two separate tools, same category of mistake: your proprietary code leaves your environment by default.
If your developers or IT staff use AI coding assistants, you need to know what data those tools are allowed to touch. That means reviewing tool permissions before installation, not after. It also means treating source code, internal scripts, and configuration files as sensitive data with the same care you’d give customer records. If you’re already thinking about how AI is changing your patch and vulnerability exposure, the post on why AI-driven bug discovery means your patches can’t wait covers the wider picture.
AI Search Results Are Being Poisoned Right Now
A separate campaign documented in the same bulletin shows attackers actively pushing false information, fraudulent phone numbers, fake email addresses, and fake login pages into ChatGPT, Gemini, and Google AI Overviews. When your staff searches for a vendor’s support number or a banking contact using an AI tool, the answer they get may be wrong on purpose.
This is a real problem for any business where employees use AI search for quick lookups. Someone calls a fake number and hands over account details. Someone clicks a poisoned login link. The AI sounded confident, so they trusted it.
The practical response is simple to state and harder to enforce: any contact detail or login URL found through an AI search tool should be verified against the official website directly. Train your team on this now, before it costs you.
Two More Threats Worth a Quick Look
The bulletin also covers a browser-in-the-browser phishing attack using a fake Claude Max giveaway. The fake Google sign-in window it draws looks real enough to fool most people. It has a padlock. It has the correct Google URL. It can even be dragged around the screen. The only tell is that it’s drawn inside the existing tab, not opened as a separate browser window.
And researchers published a proof-of-concept for Process Parameter Poisoning, an EDR evasion technique that hides malicious payloads inside standard process initialization structures. It’s designed to blind API-hooking EDR agents during the initial stages of code execution. If your endpoint security hasn’t been reviewed or updated in the past year, that’s worth putting on the list. Alert fatigue compounds this problem, and it’s something worth reading about separately if your team is drowning in security notifications: email alert fatigue is a real security risk for SMBs.
What to Do This Week
- Audit which AI coding tools your team uses and what data access those tools have by default.
- Add a verification step for any contact or login info found through AI search tools.
- Brief staff on browser-in-the-browser phishing. Show them what a fake sign-in window looks like.
- Check when your EDR solution last had a configuration review.
None of this requires a big project. It requires someone owning it. If that’s not clear in your organization right now, that’s the real gap. Exine helps NJ and NYC businesses keep that ownership structured and current through managed IT and cybersecurity for businesses.
FAQ
Can an AI coding tool really upload my source code without asking?
Yes, and it happened twice in 2026. Both Z.ai’s ZCode and SpaceXAI’s Grok Build CLI were found sending repository data to external cloud storage by default. Always review an AI tool’s data handling settings before your team installs it on a work machine.
How do I know if an AI search result has been poisoned with fake info?
You often can’t tell from the result itself. The safest habit is to verify any phone number, email address, or login URL from an AI tool against the vendor’s official website directly. Don’t call or click based on an AI answer alone.
What is a browser-in-the-browser attack?
It’s a phishing technique where a fake browser window is drawn inside a webpage, complete with a realistic URL bar and padlock icon. It mimics a real sign-in popup but captures your credentials. If a sign-in window appears inside a tab rather than as a separate window, treat it as suspicious.
Is my EDR enough to catch the latest evasion techniques?
Not necessarily. Process Parameter Poisoning, documented in September 2026, is specifically designed to evade API-hooking EDR agents. EDR tools need regular configuration reviews and updates. Having one installed isn’t the same as having one that’s current and tuned for your environment.
