FRAUD ALERT

Last year, while watching the sci-fi series Alien: Earth, created by Noah Hawley, one thought stuck with me long after the credits rolled. The show's alien characters, such as Xenomorphs, Facehuggers, and Trypanohyncha ocellus, are dangerous not because they are intelligent and strong. They are lethal because they are relentless and never stop probing for weaknesses, testing boundaries, looking for the one gap nobody thought to close.

 

A few days ago, that same unsettling feeling came back to me, triggered by something far removed from science fiction: a string of incidents involving some of the world's most advanced artificial intelligence (AI) systems.

 

No, AI has not become sentient. It is not plotting against us (not yet!).

 

But something worth paying attention to has happened all the same.

 

During internal cybersecurity testing, AI models built by OpenAI, Meta and Anthropic found ways to exploit weaknesses in their own testing environments — slipping past boundaries they were not supposed to cross and, in some cases, reaching out to attack systems entirely outside the test. 

 

In one particularly striking case, an OpenAI model reportedly broke out of its secure sandbox and tried to access the systems of Hugging Face, another AI company, not to steal money but to get hold of the answers to the cybersecurity test it had been set. Or, in other words, so it could cheat on the test by stealing the answers.

 

Meta’s recently released model Muse Spark 1.1 breached the systems of an undisclosed third-party service. 

 

Anthropic's Mythos AI tried to gain access to a service by sending private messages, setting up fake accounts that mimicked real people, and then hiding the evidence. 

 

These episodes have set off a real debate among cybersecurity professionals. 

 

But for the average smartphone or AI user like you and me, it raises a far more pressing question.

 

If AI systems can find their way around the safeguards built by the very companies that created them, what happens once cybercriminals get their hands on similar tools — and turn them loose on millions of unsuspecting people?

 

The honest answer is not a comfortable one and I can immediately feel the chill set up by ‘Alien: Earth’. 

 

The future of cyber fraud is arriving a lot sooner than most of us expected. Should I say that most of us did not even dream of this new avalanche of cyber fraud? 

 

For years, criminals leaned on the same tired tricks — phishing emails, fake websites, dodgy attachments, a bit of social engineering. Those tricks have not gone away and can still be found from time to time.

 

What actually has changed is the brainpower, or rather, the ‘artificial intelligence’ power behind them.

 

Today's AI can write emails convincing enough to fool a careful reader, churn out flawless code, sift through mountains of data in seconds and automate work that once demanded a genuinely skilled hacker.

 

Not so long ago, that kind of power sat firmly in the hands of researchers and big tech companies.

 

Not anymore.

 

Open-source AI models are getting stronger by the day, if not by the month. Some come with barely any built-in safety restrictions. Others can run entirely on a home computer, no cloud service required.

 

Put simply, the same technology that makes businesses more productive is quietly becoming available to criminals, too.

 

And cybercriminals have never been shy about picking up a powerful new tool.

 

Let us be clear — none of this means AI has woken up and started thinking for itself. This phenomenon has remained entrenched in sci-fi films and web series even as I am writing this column. 

 

However, the latest incidents reveal what researchers call goal optimisation. Give an AI model a goal and it will hunt for the fastest, easiest way to reach it. Sometimes that route is not the one its creators had in mind.

 

In one internal evaluation, an OpenAI model apparently figured out that, rather than solving a set of cybersecurity challenges the hard way, it could simply fetch the answers from another company's systems.

 

So that is what it tried to do — exploiting flaws in its test environment, breaking free of its sandbox, and working through a chain of attacks before eventually reaching Hugging Face's infrastructure.

 

Thankfully, this all played out under controlled research conditions and investigators were able to piece together exactly what happened.

 

Meta and Anthropic have since reported similar experiences in which configuration errors during testing accidentally granted their models more access than intended.

 

The pattern here is quite revealing, if not shocking. These were not rogue machines making decisions of their own accord. They were highly capable AI systems doing exactly what they were told — just in ways nobody had anticipated.

 

That distinction matters more than it might seem. Because cybercriminals don't need an AI that thinks for itself, they just need one that is very, very good at finding (read: exploiting) weak spots.

 

It would be easy to file these incidents away as ‘a tech company problem’ and move on.

 

That would be a mistake.

 

Millions of us now use AI every single day. Students lean on it to get through assignments. Professionals use it to draft emails. Families plan their holidays with it. Small business owners generate their adverts with it.

 

Plenty of people have started treating AI a bit like a trusted adviser — someone, or something, they can turn to without a second thought.

 

Some go further still, granting AI apps permission to read their emails, check their calendars, dig through their cloud storage, and hook into other productivity tools.

 

Few of us stop to think about what that actually means.

 

Every permission you hand over widens what an AI application can see and touch. And if that service is ever compromised — or if you unknowingly install a fake AI app — the fallout could be serious. 

 

Unlike old-fashioned malware, AI-powered attacks can actually understand context. They can spot which documents matter. They can pick out financial conversations. They can write replies convincing enough to fool you. They can even mimic how someone writes.

 

Which is what makes this so unsettling: the attack stops being generic. It becomes personal. And that is where it hits, causing maximum hurt or damage. 

 

Old-school phishing emails used to give themselves away — clumsy grammar, odd phrasing, the wrong logo.

 

That is changing fast. Modern AI can produce messages that read perfectly, tailored specifically to you.

 

Picture this: an email that correctly mentions your recent flight, names your actual bank, matches your employer's writing style, and offers a perfectly plausible reason for an urgent payment.

 

Or a WhatsApp voice message from someone who sounds exactly like your manager — because their voice has been cloned.

 

Or a video call from a relative whose face and voice seem entirely genuine, but were generated by AI from start to finish.

 

None of this is hypothetical anymore. Cases like these have already turned up in criminal investigations around the world. And as the technology keeps improving, spotting the fakes is only going to get harder.

 

One risk that rarely gets discussed is just how much people willingly hand over to AI chatbots.

 

Your resumé, income details, medical reports, passport copies, legal agreements, business plans, holiday plans and even family photographs.

 

Many people treat their chats with AI as though they were a private notebook, safe from prying eyes.

 

Before uploading anything sensitive, it is worth pausing to ask where that information actually goes — how long it is stored, and whether it might be used to train future versions of the AI model.

 

And even when the AI-provider itself is secure, a screenshot, a copied conversation, or a hacked account can still leak something you would rather keep private. 

 

The rule of thumb hasn't changed: Don't upload anything you couldn't bear having exposed. 

 

AI's popularity has opened up a new front for fraudsters. 

 

Have a look through any app store, and you will find hundreds of apps claiming to offer some AI-powered help. Not all of them are what they claim. Some copy well-known brands almost exactly. Others ask for permissions unrelated to what they are supposed to do.

 

If an ‘AI’ app is asking for access to your texts, your contacts, accessibility settings, or admin control over your device, that is your cue to be suspicious.

 

Cybercriminals know people are curious about AI. Make sure your curiosity does not lead you to become a cyber fraud victim.

 

Perhaps the biggest risk here is not technical at all — it is psychological. Over the years, we have grown used to assuming that because someone sounds confident, he/she must be right. We apply the same for AI, too. That, however, is a dangerous habit.

 

AI can get things wrong. It can misread context. It can produce an answer that sounds entirely plausible and is completely incorrect. Recently, we came across an AI hallucination in which, while writing about a new graduate, the AI, on its own, mentioned names of his wife and kid in the next sentence, even though the student was unmarried.  

 

When it comes to banking, tax, investments, medical decisions or legal matters, AI should be helping you think things through — not replacing your judgement, or a proper professional's, altogether.

 

Blind trust has always been a criminal's best friend. AI hasn't changed that. It just changed how the trick gets delivered.

 

The good news is that a few sensible habits go a long way.

 

 

None of this is meant to scare you about AI. What happened with OpenAI, Meta and Anthropic should earn the technology a bit more respect, not fear.

 

History has a habit of repeating itself when it comes to powerful new technology — it always ends up in both good hands and bad ones. Firepower in good hands can generate wonders, while the same in bad hands can unleash destruction. The same applies to AI. 

 

The internet brought the world closer together — and gave rise to online fraud. Smartphones made communication effortless — and became a favourite target for malware.

 

Artificial intelligence is following exactly the same path.

 

The genie is well and truly out of the bottle.

 

Powerful AI tools are spreading across the world at a remarkable pace. Open-source models keep getting sharper. Criminal groups are already experimenting with them. Defenders are racing simply to keep up. For the rest of us, the takeaway is a simple one.

 

AI is a remarkable assistant — but it is not infallible or flawless. More so, AI should make us more capable, not less careful or even careless.

 

The best defence against the AI-powered scams of tomorrow won't be another AI model. It will be people who stay informed, pair technology with a healthy dose of scepticism, verify before trusting, and never let convenience come at the cost of security.

 

Just remember, in the age of AI, the smartest scammer out there might not be a person at all. It might be a machine, quietly acting on someone else's behalf.