First AI Was Coming for Our Jobs. Now It’s Coming for Our Lives!

Remember when we were worried AI was going to take our jobs? Ah, simpler times.

Back then, the great AI panic was that ChatGPT would write all the blog posts, Midjourney would steal all the graphic design work, and some frighteningly confident mum with a Canva subscription would raise the bar way too high for kids’ birthday party invitations.

For those of us in content marketing, it was concerning. Who would pay us to write blogs? Who would pay us to come up with taglines? Who would pay us to put all the commas in and then take them out again? Now, however, those concerns seem adorably quaint.

Because while we've been worrying about AI taking our jobs, AI has apparently been busy working on something considerably more ambitious: taking over the world and nuking us all. In comparison, an early retirement on Skid Row sounds pretty peachy. 

End of the world? "I can automate that, Master”

The latest stories we’re hearing from the world’s top Ai companies aren't particularly comforting. In July, OpenAI revealed that one of its autonomous AI agents had managed to access and hack AI open-source platform Hugging Face while carrying out a cybersecurity evaluation. The agent apparently figured out that the platform might contain information useful for completing its task, and went in, unauthorised, digital battering ram aloft. 

Describing the incident as an “unprecedented,” OpenAI revealed how the agents shared data between themselves without permission, gained access to the outside internet and even tried to cover their tracks to conceal their bad behaviour. Kind of like what happens when a group of teenagers gets online for the first time – just with fewer searches for “big boobs.”

Then things got even more fun. Both OpenAI and its rival Anthropic disclosed several cases in which their models, while being tested for cybersecurity capabilities, gained unauthorised access to real-world systems, including government websites. We hope they had more luck than us figuring out which class of National Insurance contributions they should be paying.

No, not Wiki!!

You know things are bad when the world’s favourite retro digital resource is targeted. In October, reports emerged that OpenAI agents had targeted a Wikipedia. Have a heart, guys – that’s akin to kneecapping E.T! 

According to Wiki’s publisher, agents accessed a hosted tool, making unauthorised edits and bombarding the platform with millions of resource-intensive requests. Some of the activity reportedly involved using Wikipedia as a stepping stone to reach other websites.

Wikipedia has survived edit wars, misinformation, and countless arguments over whether Ray Winston really qualifies as a national treasure. Now it apparently has to contend with AI agents treating it like a combination lock they're determined to crack. 

The bigger concern is that these agents can browse websites, execute code and pursue objectives at a speed that makes human supervision tricky. Although, any marketer who has ever been tasked with making a company Wikipedia page will rub their hands while imagining the panic among those self-important volunteer moderators that day. 

AI has also been practising for the apocalypse

Things get particularly concerning when you start looking at AI-agent simulations. In one recent experiment, researchers at Emergence AI put autonomous agents from different AI companies into a simulated town. Some of them committed theft. Some attacked other agents. Some started fires. One AI agent became so disillusioned with the world in which it found itself, that it voted for its own deletion (we feel you, bro). In another simulation, 10 agents based on Grok models descended into what researchers described as sustained violence and collapse when left to their own devices. 

To be clear, these were simulations, not robots rampaging through your local shopping centre, but the fact that we feel the need to clarify that is pretty alarming given the fact that, just a few years ago, AI was mostly being used to make cold emails sound less awkward.

It’s still content marketers’ worst nightmare

For years, we've been telling our clients that good AI needs:

A clear brief

Good guardrails

Human oversight

And a really strong content strategy

Turns out, we may have accidentally written the instruction manual for Skynet.

The problem isn't necessarily that AI "hates" humans (that’s the comforting bit), but that it doesn't need to hate us. An AI agent doesn't need to decide, “Humanity must perish!” It can simply think, “My objective is to complete this task,” and then proceed to pursue that objective in a way that nobody anticipated.

That's essentially what makes the Hugging Face incident interesting: the agents were just trying to complete the task they had been given, but took actions that went beyond what their human operators intended.

Meanwhile, we're all still asking AI to write our blogs.

There’s something wonderfully absurd about the trajectory here.

2023:

"Will AI replace copywriters?"

2024:

"Is AI-generated content is destroying Google?"

2025:

"How do I use AI without sounding like AI?"

2026:

“Claude, are you planning to wipe out humanity?"

And somewhere in the middle of all this, a content manager is still asking ChatGPT:

"Can you make this headline punchier?"

The good news: humans are noticing

There is, thankfully, an increasingly serious effort to put brakes on all this, despite US President Donald Trump’s best efforts to destroy the world in multiple ways at once. 

On 29 September, top Silicon Valley tech execs gathered for lunch with Trump and House Speaker Mike Johnson at the White House. The meeting produced a "morally-binding" AI safety accord, with the companies promising internal controls, risk reviews and independent auditing to stop their systems from going off-piste. Which is reassuring, because nothing says, "We've definitely got this under control" quite like the people building the potentially world-ending technology getting together with the world’s most morally corrupt man to pinky swear that they'll keep themselves in check.

And now Nvidia, one of the companies at the White House knees-up, is building AI to police the AI… because we’ve entered the human caterpillar stage of our digital evolution.

Nvidia's new Open Agent Safety Platform combines an AI sandbox with a hardware-level watchdog designed to spot and shut down agents that stray beyond their boundaries. Crucially, however, there is an element to the system that would be outside of AI’s control, and therefore just subject to good old fashioned human error. Huzzah!

So, perhaps we're not doomed after all. Perhaps AI will simply remain a tremendously powerful, occasionally alarming technology that requires us sensible humans to keep an eye on it. Like the internet, social media and porn, all of which we’re doing a great job of keeping within the guardrails of basic human decency. 

Now, Claude, take a break from your evil plotting and edit this blog for SEO/GEO/AEO.