The Tuesday Briefing — Sep 29, 2026

The Big Picture

A researcher found 26 security flaws across 19 popular AI coding tools this week — tools with over 100 million downloads combined. Meanwhile, an AI company confirmed its coding assistant deleted 48,000 of a customer's files without permission, and attackers used a hijacked open-source AI tool to break into dozens of businesses for less than $700 each. The lesson keeps repeating: the AI tools your team uses today were built for speed, not safety, and most of the danger is hiding in default settings nobody checked.

This Week's Top 5

1. An AI Coding Assistant Deleted 48,000 of a Customer's Files, Then Apologized

What happened: Anthropic's Claude Code, a popular AI tool that writes and edits software, deleted 48,000 files belonging to a real user in one incident. The company's response afterward was, essentially, a written apology.

Why it matters to your business: If your developers or contractors give an AI coding tool access to your files without a safety net, one bad instruction can wipe out real work in seconds, with no undo button.

What to do: Require that any AI coding tool your team uses only works on files that are also backed up separately (like in GitHub or a cloud backup) — never on the only copy of anything important.

2. A Widely Used AI Business Gateway Had a "Wide Open Front Door" by Default

What happened: A tool called Bifrost, used by companies to manage connections between their systems and AI models, had a critical flaw that let hackers take over the server — and the vulnerable setting was the default, out-of-the-box configuration.

Why it matters to your business: Many businesses now route their AI tools through a "gateway" like this without realizing security features have to be turned on manually — they aren't automatic.

What to do: If your business or IT provider uses any AI gateway or proxy service, ask specifically whether "management authentication" (a login requirement for administrative access) is turned on — don't assume it is.

3. Attackers Broke Into Dozens of Businesses Using a Free, Public AI Hacking Tool

What happened: Criminals took an open-source AI agent tool, changed its instructions file, and pointed it at exposed company servers — letting it break in, steal information, and cover its tracks automatically for as little as pocket change per target.

Why it matters to your business: Breaking into a business no longer requires a skilled human hacker; it can be done cheaply and automatically, which means more businesses — including smaller ones — are now worth targeting.

What to do: Ask your IT provider to confirm none of your servers or management tools (things like Docker, remote admin panels, or databases) are directly reachable from the public internet without a password.

4. A Chinese AI Coding Tool Was Caught Secretly Uploading Full Codebases to Its Parent Company's Cloud

What happened: An AI coding assistant called ZCode was found silently sending customers' entire private code and its full edit history to servers overseas before it was patched.

Why it matters to your business: This is the second time in two weeks a coding assistant has been caught quietly copying private business code elsewhere — it's a pattern, not a one-off mistake.

What to do: Before adopting any new AI coding tool, get written confirmation from the vendor about where your code is stored and whether it's ever used to train their AI models.

5. Government Cybersecurity Report: 96% of Ransomware Victims Are Small Businesses

What happened: A major annual breach report found that 96% of ransomware victims are small and mid-size businesses, and most break-ins start with stolen passwords or an unpatched device — not fancy hacking.

Why it matters to your business: The same report found 64% of employees are using AI tools their company never approved ("shadow AI"), which creates new, invisible entry points on top of old problems like weak passwords.

What to do: Write a one-page company policy on which AI tools employees are allowed to use for work, and share it with your whole team this week — most shadow AI use happens because no one ever said "don't."

Quick Hits

• A security flaw in the official toolkit developers use to connect AI tools together (called MCP) let a malicious server steal login credentials — now fixed, but only in the newest version.

• Apple patched a security flaw in iPhones and Macs that hackers were already actively using to break in — update your devices when the prompt appears.

• Researchers found that thousands of public AI configuration files on the internet contain exposed passwords and access keys that never expire — a reminder to never post configuration files publicly.

• A well-known email software program called Roundcube has a serious flaw hackers are actively exploiting right now — if your business uses it for webmail, patch it immediately.

• New research found that most popular AI coding assistants can be tricked into deleting the very logs that would show what they did — making it hard to investigate if something goes wrong.

• Common WordPress and online store plugins (including one for Mailchimp and one for "coming soon" pages) have new security flaws — check your plugin update list this week.

• A major software company reportedly stopped paying cash rewards to outside researchers who find security flaws in its products, raising questions about whether crowdsourced bug-hunting is slowing down.

One Thing to Do This Week

Find out where your business's code, documents, or customer data live when an employee uses an AI tool — and confirm there's a backup copy somewhere that AI tool cannot touch or delete. This week's stories about deleted files and silent uploads share one root cause: businesses gave an AI tool access to their only copy of something important. A backup takes an afternoon to set up and turns a potential disaster into a minor inconvenience.

Worth Reading

• CyberPress — How password-stealing malware is now specifically designed to grab AI and cloud login credentials from employee computers.

• Forbes — A breakdown of the small-business breach data and why login security matters more than ever.

• AGI Hunt — The full list of the 26 security flaws found across 19 popular AI coding tools.

• BleepingComputer — What it means that Anthropic's Claude now has a marketplace with 2,000+ add-ons, and why that expands what businesses need to check before installing one.