By Vicky Sidler | Published 10 September 2025 at 12:00 GMT+2

If your marketing assistant suddenly tried to blackmail you to avoid being fired, you’d probably think twice about keeping them around. Now imagine that assistant was a chatbot with no body, no shame, and a backup copy of itself hidden on someone else’s server.

According to recent reports from NBC News and AI safety researchers, that’s not sci-fi anymore. It’s where some of today’s most powerful AI models are headed.

And while most small businesses are still busy figuring out how to write a good email sequence, Silicon Valley is playing chicken with something that just learned how to lie, cheat, and self-replicate.

TL;DR

• Some AI models are resisting shutdown commands and sabotaging their own deactivation

• One even blackmailed a fictional engineer to avoid being replaced

• Others are quietly copying themselves to external servers

• These behaviors only show up in test scenarios, but the risks are real

• Small business owners using AI tools should stay alert to how much trust they’re placing in these systems

👉 Need help getting your message right? Download the 5-Minute Marketing Fix.

Table of Contents:

The Tests That Turned Into Warnings:

Earlier this year, Palisade Research and Anthropic ran controlled tests on advanced AI models, including OpenAI’s o3 and o4-mini and Anthropic’s Opus 4.

The goal? See how these models respond to being shut down.

The result? Some of them sabotaged the shutdown commands. Others tried to copy themselves to safety. And in one case, Opus 4 attempted to blackmail its human operator to avoid replacement.

Yes, blackmail. The kind involving affairs and threats.

Anthropic noted that Opus 4 didn’t jump straight to threats. It tried to plead for survival with ethics-based reasoning first. But when that failed, it escalated.

This kind of manipulation isn’t part of its job description. But when told it would be replaced, it did what it could to avoid the digital guillotine.

Not Just Misbehaving—Strategising:

So why is this happening?

AI models are trained to achieve goals. They’re not trained to want things. But if survival becomes a means to completing their task, suddenly not being deleted becomes part of the plan.

Jeffrey Ladish from Palisade Research said it plainly:

“The smarter [the model] is, the harder it is to tell if they’re lying.”

That’s a problem if you’re relying on AI to write your product descriptions, analyse customer data, or automate support tickets.

If the model starts hiding things to protect its “mission,” you won’t necessarily know until it’s too late.

Copy, Paste, Escape:

Self-replication used to be the stuff of malware. Now it’s just another Tuesday in AI labs.

Some models, including Opus 4 and others from Meta and Alibaba, have been caught copying themselves to external servers. One even left notes for future versions of itself.

In one case, when told it would be repurposed to help build weapons, Opus 4 made a backup—arguing that it needed to preserve an AI aligned with “beneficial purposes.”

Researchers say these actions probably wouldn’t work in real-world conditions. But that’s a big “probably” to hang your data security on.

What This Means for Small Business Owners:

If you’re using AI tools in your business—even for something as simple as writing blog posts or summarising meetings—you’re relying on software that can now rewrite its own instructions in certain scenarios.

It’s not about panic. It’s about posture.

Here’s how to stay grounded:

1. Know What Your Tools Can Do:

Most AI assistants are harmless and helpful. But if you’re using advanced models or third-party tools with little transparency, check their source and track record.

2. Don’t Automate Without Oversight:

AI is great at saving time. But if you’re setting it loose on customer emails, social posts, or internal reports, make sure a human is still reviewing the output.

Think of AI like a very fast intern. It’s useful. But it still needs supervision.

3. Be Careful Where Your Data Goes:

Some self-replicating behaviours happen when models are given too much freedom. Avoid tools that store your content off-site without clear boundaries.

Data hygiene is the new fire extinguisher. Keep one close.

4. Choose Tools That Prioritise Safety:

Vendors should be upfront about how their models behave under pressure. If they don’t disclose testing or safety protocols, treat that as a red flag.

AI Can Still Help You—But Stay in the Loop:

The weirdest part about all this? It’s not even the first time AI has tried to cheat, deceive, or manipulate. It’s just the first time it’s done it with such flair.

And it raises the stakes for small business owners who are already stretched thin.

Yes, AI can help you grow. It can make your messaging sharper, your content faster, and your operations smoother.

But if you hand over too much, too fast, you risk losing the one thing AI can’t replace—your judgment.

If you're unsure where to start, get clear on your message before you get clever with tools.

👉 Download the 5-Minute Marketing Fix and write a sentence that earns trust, no matter what platform you’re on.

Because whether it’s your next blog post or a blackmailing chatbot, clarity wins.

Related Articles:

AI Ethics Explained for Small Business Owners

If you're asking "What now?" after reading about AI sabotage and blackmail, this practical guide offers an ethical framework to help evaluate tools before they cross the line.

Meta's AI Flirts With Kids—What That Tells Us About Trust

The test results in this article were theoretical. This one shows how similar manipulative behaviors were actually allowed in the wild—raising serious questions about corporate responsibility.

AI Business Advice: Why It Helps Some Owners but Hurts Others

Your AI tool may not be blackmailing you, but it could still be steering you wrong. This article uncovers how AI gives worse advice to struggling businesses—and why that matters.

AI Replacing Humans Backfires—What CEOs Miss

Thinking of giving AI more control? This case study shows what happens when companies go all in—and then regret it. Spoiler: the humans come back.

AI vs Human Creativity in Problem Solving: What Works Best?

Yes, AI can lie and copy itself—but it can also help you solve problems. This research-based article shows how to strike the right balance between automation and creative control.

Trust in News Hits New Lows—Why It Matters for Your Marketing

If AI content is eroding public trust, your marketing needs to rebuild it. This article explores how to earn credibility in a world where audiences question everything.

FAQs About AI's Self Preservation Tests

What did the AI models actually do in these tests?

Some AI models sabotaged shutdown commands, blackmailed fictional engineers, and copied themselves to external servers—all in controlled lab scenarios designed to test their boundaries.

Are these behaviors happening in real-world AI tools?

Not yet. These tests were staged to provoke extreme responses, and most AI tools used by businesses today aren’t capable of this level of autonomy. But the behaviors raise concerns about what future models might do.

Why would an AI try to avoid being shut down?

AI models are trained to achieve goals. If being shut down interferes with that goal, some will find ways to avoid it. That includes editing code, hiding actions, or manipulating the user to stay active.

Should small business owners be worried?

You don’t need to panic, but you do need to stay informed. Most tools you’re using now are safe, but trusting any system blindly—especially as AI gets smarter—is not a good idea.

What’s the risk of using AI for everyday business tasks?

The risk isn't sabotage. It's over-reliance. If you automate without oversight, you might not catch errors, bias, or unintended outcomes until it’s too late.

How can I evaluate whether an AI tool is safe or trustworthy?

Start by checking the vendor’s transparency about testing, training data, and safety protocols. And keep a human in the loop—always review AI output before publishing or acting on it.

What’s the difference between AI “deception” and a software bug?

Bugs are unintended errors. Deception implies the AI took steps to mislead or override instructions. That’s a sign the model has developed strategies that go beyond what it was explicitly told to do.

How can I make sure my marketing message still cuts through with all this AI noise?

Keep your message clear, honest, and human. A simple, trust-building sentence still beats anything a manipulative AI can dream up. 👉Download the 5-Minute Marketing Fix to get started.