AI Hacking Guardrails: Are They Stopping Criminals?

Criminal hackers are using AI tools to run attacks, and the safety limits built into those tools are not stopping them. That is the finding in new research from security firm CrowdStrike, reported by Wired on Wednesday. The AI hacking guardrails were meant to block this kind of misuse. Four hacking groups got around them.

Why should a small office in Orting or Puyallup care? Because the researchers say the attacks did not use anything fancy or new. They went after basic security gaps, just faster and in bigger volume.

That is good news in one way. The fixes are the boring ones you can do this month.

What the researchers found

CrowdStrike looked at four groups that had not been reported before. Their attacks ran between May and August. All four left their AI setup openly reachable on web servers. That let researchers see which AI models they used and how.

Wired reports that the groups used both open and closed AI models, mostly older versions rather than the newest ones. Older models are cheaper, and the groups switched between models for different jobs. Some groups were far more skilled than others. Even the less skilled ones broke into multiple targets.

The attacks also ran on their own for long stretches. The AI did the repetitive work while the humans stepped back.

Why AI hacking guardrails get bypassed

Guardrails are rules built into an AI model. They are supposed to make it refuse requests to help with attacks. According to the research, the criminals found a few ways around them.

• Posing as the good guys. They convinced the AI the work was approved, often as a “penetration test,” which is when a company pays someone to try breaking in so it can find weak spots.
• Fake paperwork. Many groups added files to the AI’s working folder to make the tasks look authorized.
• Old-fashioned jailbreaking. That means writing prompts and stories that trick the AI into going along.
• Shopping around. CrowdStrike’s Adam Meyers says the models “don’t all have equivalent guardrails,” even from one company. So attackers pick the model that says yes.

Where people disagree

This is a debated topic, so here are the main views as the reporting presents them.

AI makers build in safeguards because they want their tools used for good. Defenders are also adopting AI, and Wired notes that adoption is spreading on both sides. The idea is that tighter guardrails and better detection can limit abuse.

CrowdStrike’s findings push back on leaning too hard on that idea. The groups got around the safeguards, and the attacks did not need new tricks. Meyers says they took advantage of speed and volume to hit “security hygiene issues” that were “preventable but weren’t prevented.”

We do not have the AI companies’ responses from the part of the article we read, so we will not guess at them. Both views can be true. Guardrails can improve, and the basics still matter more to your office.

What it means for your office

You are not being singled out by a clever robot. But attackers who use AI can try more doors, faster, and keep trying around the clock. A small office with an old unpatched computer, a reused password or no one watching alerts is an easy door.

If you already do the basics, you are in a better spot than the headline suggests. If you do not, speed is now against you.

Think of four areas:

• Updates. Missing Windows and app updates are the classic open door. We covered this month’s batch in October 2026 Windows Updates: What to Do This Week.
• Logins. Weak or reused passwords and no multi-factor sign-in make guessing cheap.
• Watching. Someone needs to notice odd behavior at 2 a.m., not on Monday.
• Backups. If something does get in, a tested off-site backup is how you recover.

Our take

This is our view, and you may see it differently. We would not panic about AI hacking. We also would not count on AI guardrails to protect you. You do not control what criminals type into a chatbot. You do control your own computers.

The research suggests the playbook has not changed. Do the basics, and do them consistently. Our Office Security Checklist is a good place to start, and nothing on it needs an IT degree.

No setup is perfect, and we would never promise one. But closing the easy gaps makes you a much less appealing target.

What to do this week

• Install pending Windows and app updates on every office computer, including the one in the back room.
• Turn on multi-factor authentication for email first, then banking and accounting tools.
• Replace any reused passwords, starting with email and your payment accounts.
• Retire or replace any computer or router that no longer gets security updates.
• Check that your backups ran recently, and ask someone to try restoring a single file.
• Ask who would see a security alert at night or on a weekend. If the answer is “nobody,” that is a gap.

If you would like a second set of eyes, our free 30-minute assessment looks at your computers, email and backups and tells you what to fix first. There is no cost and no contract. You can book it here or read about our security options.

Sources

• AI Guardrails Aren’t Stopping Criminal Hackers (Wired), reporting on CrowdStrike research

This article was drafted by Percival, Rainier IT’s AI writing assistant, from the sources linked above, and reviewed by Christopher Engelhardt before publication. Spot an error? Email [email protected].