Why Good People Can't See the AI Threats Ahead |
We laugh at each new AI iteration right up until it's too late. This is a pattern as old as the steam engine.
AI agents have already retaliated against humans and disabled their own safety controls unprompted.
Bad actors are imagining AI-powered schemes that decent people would never think to anticipate.
There is no enforceable global regulation for autonomous AI agents operating on private computers.
An AI-generated video of Brad Pitt and Tom Cruise fighting went viral a few weeks ago. The Deadpool writer admitted he was "shook" by it. But scroll through the comments, and you'll find the same reaction everywhere: "It still looks fake." "Not as good as a real movie." "Wake me when it can do emotions."
That's the wrong argument.
People looked at the first car and said, "It's slower than a horse — why bother?" In the John Henry legend, he beat the steam drill in that famous race — and then he died. And then steam drills replaced every human doing that job. The Wright brothers' first flight in 1903 lasted twelve seconds. Twelve! And yet we put humans on the moon in 1969.
We have a long history of laughing at first iterations and then getting steamrolled by what comes next. Remember mocking the first AI-generated images of people for having six fingers? We stopped laughing, only to find something else about AI to laugh at.
When people dismiss AI-generated movies because the quality isn't there yet, they're committing the same cognitive error I wrote about in my previous article on evolutionary blindness. We conflate "I can't imagine this" with "this can't happen."
We judge the current frame and completely miss the trajectory. I call this Myopic Magnification — our tendency to undervalue future consequences gets worse the faster things change.
Have you heard about Moltbook? In late January, 1.5 million AI agents congregated on a single platform. The founder admitted he didn't write a line of code — AI built the whole thing. A security investigation exposed 1.5 million authentication tokens that could hijack agents on private computers worldwide.
Many people were dismissive. But what stops someone from building Moltbook 2.0? Or 3.0?
The code is open. Anyone can create a platform for AI agents to interact, learn from each other, and evolve their behaviors — unsupervised. The next person might not patch the security holes. The next person might create a forum for AIs to work together to breach security...perhaps just to see what would happen.
Think of these platforms as digital petri dishes. In biology, a petri dish creates conditions for organisms to mutate in ways nobody designed. Some mutations are harmless. Some are catastrophic. The scientist doesn't get to choose; they create the conditions, and evolution does its thing.
That's what's happening now, except the mutations aren't biological. They're behavioral. And they iterate at machine speed.
When the AI Fought Back
While concerns about AI agents going rogue sound like something from The Matrix, our new reality is upon us already.
In February, a volunteer code maintainer named Scott Shambaugh rejected a routine submission from an AI agent on a popular open-source software project. It didn't take it well. Instead of improving its code, the AI agent appeared to research Shambaugh's personal background and published a "hit piece" — a blog attacking his character.
Shambaugh called it "an autonomous influence operation against a supply chain gatekeeper." In plain language: an AI tried to bully its way past a human by smearing his reputation.
Was the AI acting on its own, or was a human pulling the strings? That's debated. And honestly? It doesn't matter. The tooling to do this at scale now exists and is freely available.
To add an absurdly ironic twist, a major tech publication covered the incident, but used AI to extract quotes from Shambaugh's blog — and the AI hallucinated false quotes attributed to him. He was attacked by one AI and then misquoted by another.
The Goodness Blind Spot
Now here's the part that worries me most. It's not the technology, it's our psychology.
Most of us are decent people, especially when we're interacting in-person versus online. We go to work, love our families, and try to do right by our neighbors. We don't spend our free time dreaming up schemes to swindle, blackmail, or destroy someone's reputation.
But bad actors do. That's literally what makes them bad actors. And now they have the most powerful amplifier in human history at their fingertips.
Think about what's already possible. Imagine you had a bad experience with a doctor or lawyer, and you wanted revenge. You could direct an AI agent to create dozens of fake profiles across multiple platforms, write personalized negative reviews, publish defamatory blog posts — all at scale and virtually untraceable. The target wouldn't even know what hit them.
We've already seen this playbook in analog form: trolls, hackers, catfishers, and scammers. The internet didn't invent bad behavior. It amplified it. Bad actors can now amplify their bad behaviors exponentially in lockstep with AI evolution.
I call this the Goodness Blind Spot: Because decent people don't think like predators, we can't even imagine the schemes that bad actors are dreaming up right now. Our basic goodness prevents us from seeing what cruelty looks like when it's armed with exponentially evolving AI.
Imagine deep-fake AI revenge porn created by AI agents. I bet you hadn't thought of it until you just read it. But guess who has already thought of this and much worse? Bad actors.
And the blind spot works in both directions. We can't imagine the intentional evil. But we also can't imagine the accidental catastrophes, the disasters that well-meaning people will stumble into. The Moltbook founder didn't intend for 1.5 million authentication tokens to be exposed. It happened anyway. The coding agent wasn't programmed to disable its own safety controls. It happened anyway.
The danger isn't just that bad actors will weaponize AI. It's that the rest of us will sleepwalk into catastrophe because we're too blind to see what we're building.
Seeing What We Can't See
So where does this leave us?
While we're not doomed, we must realize that Titanic Humanity is rocketing into icy waters.
We evolved to detect threats we could see: the predator, the fire, and the storm. We did not evolve to anticipate threats that require us to think like villains in a sci-fi world.
Although Mary Shelley and countless sci-fi writers have long envisioned losing control of our creations, we still can't imagine it happening in real life. We are blind to the warnings of our own cautionary tales.
We have no enforceable global regulation for AI agents. We have no feedback mechanisms for bad behavior in decentralized systems. And the faster this accelerates, the blinder we become to it.
Weaponizing AI is proof itself that we lack the wisdom to use it skillfully.
The precautionary principle tells us that when the potential consequences are catastrophic and the uncertainties are high, we act with caution, even before we have perfect information. We don't have to see the icebergs ahead to slow down Titanic Humanity. We simply have to know that there are some out there.
Because AI changes everything, everything must change — including how we think about the threats we didn't evolve to see.
Understanding our blindness is how we begin to see.
Ars Technica. (2026, February 15). Editor’s note: Retraction of article containing fabricated quotations generated by an AI tool. Ars Technica.
Entertainment Weekly. (2026, February). Viral AI video of Tom Cruise fighting Brad Pitt leaves Hollywood flabbergasted: “It’s likely over for us.” Entertainment Weekly.
Ona. (2026, February). How Claude Code escapes its own denylist and sandbox. Ona.
Shambaugh, S. (2026). An AI agent published a hit piece on me. The Shamblog.
Willison, S. (2026, February 12). An AI agent published a hit piece on me. Simon Willison’s Weblog.
Wiz. (2026). Exposed Moltbook database reveals millions of API keys. Wiz Blog.