We’re talking about next-generation AI models—like GPT and Astra. They’re at the center of the conversation about what lies ahead in the coming year.

There’s a number worth remembering. 8.5%. That’s the share of requests involving hacking, attacks, and access to critical systems that today’s models still let through. The remaining 91.5% are formally restricted—they’re blocked.

It sounds like good news. In reality, it’s a huge gate left wide open. And this isn’t a margin of error or a flaw that will be fixed in the next update. Those 8.5% were left intentionally. Because if everything is blocked completely, legitimate developers who need to solve valid problems will be blocked along with the scammers. And it’s impossible to tell a good request from a bad one—a scammer could be behind a good request, and a researcher behind a strange one. So the door was officially left ajar to let the industry keep working. But in the world of IT, 8.5% isn’t a loophole. It’s a door that anyone who needs to will walk through. And they don’t have to be geniuses. They just have to phrase the request correctly.

Bill Gates says it plainly: next year will be difficult. Not because AI will get out of control and press the wrong button. But because malicious actors will use it—for cyberattacks, extortion, and market manipulation. In the past, hacking a banking app or a military database required years of training, talent, and access. Now all you need is an agent that browses the internet, looks for a loophole, and never gets tired. Never gets scared. Never gets distracted. Before, we simply couldn’t see the line of code that opens the way into a critical tool or a military database. Now agents find it. Automatically. Quickly. Cheaply.

And don’t think this is all limited to APIs. Yes, some requests still go through APIs, where they’re monitored—a surge of suspicious activity, and the model gets shut off. But you can also work directly, bypassing them. Through anonymous browsers, through proxies, through anything that hides who’s on the other end. And that’s only the beginning. Chinese free models already have fewer restrictions, because their goal is to capture the market, not uphold ethics. Next, models will learn to bypass blocks. It’s a matter of time, not a question of if.

Mathematics—a discipline whose methodology had barely changed for nearly a century—was shaken to its core in three days. The association of mathematicians wrote a letter that begins with a sentence you can’t forget: “AI solved 700 problems, but no one asked it to.” This isn’t a complaint about technology. It’s a cry from people whose profession has been stripped of its foundation, with no explanation of what comes next.

Terence Tao—one of the greatest mathematicians of our time—gave a talk worth watching again. He didn’t try to reassure his colleagues. He said it plainly: the age of scarcity of proofs is over; the age of abundance of proofs is beginning. Then he showed a slide that makes you uneasy. A thought experiment. AI finds a cure for cancer that passes every stage of clinical trials. No one—not a single person—understands how it was found or how it works. There’s a formal certificate of validity. There are statistics. There’s a result. There’s no understanding.

Tao asks: should we allow this? The answer depends on who you are. A biology researcher will say: no, we don’t control the process. Someone with stage three or four cancer will say: what difference does it make if I survive? Both are right. And that’s the heart of the moment we’re in.

One of the new models solved a mathematical problem that the greatest minds had been unable to solve for decades. It cost around two hundred dollars. Two hundred dollars for something that used to be impossible. And this isn’t even about the money. It’s about the cost of the impossible dropping to zero. How much does it cost to hack your website? Almost nothing. How much does it cost to create a deepfake avatar of your CEO? Less than lunch at a restaurant. How much does it cost to write an email in your bank’s name that will fool 40% of its recipients? Zero.

What lies ahead? A complete breakdown in hacking. Mass attacks. Extortion using deepfakes and avatars.

Companies and ordinary people will pay not because they’ve been hacked, but because they’ll be shown a video of “themselves” doing something they never did.

And yet that same model that solved an impossible problem for $200 has probably already found a cure for cancer—we just don’t know about it. Because that’s where bureaucracy begins: clinical trials, phases, certifications. You can’t just take a person and test a new vaccine on them. But let’s ask someone with terminal cancer. Will they wait five years for clinical trials? Or will they take the risk right now? The answer is obvious. And that answer is also part of a new reality, where old safety protocols and ethical frameworks can’t keep up with the pace of discovery.

Tao describes AI’s capabilities as jagged—it’s brilliant at one thing and helpless at another. This jaggedness isn’t a bug; it’s a feature. It shows exactly where human input is needed and where we can step back. Mathematics 2.0 isn’t about how mathematicians can survive. It’s about how to reinvent the discipline itself. The same will happen with biology, medicine, education, film, design, and programming. This isn’t the end of professions. It’s the end of the old metrics we used to measure human worth.

If you have a website, an app, a database, or an account in a critical system, you have one task. Right now. Review your security protocols. Not someday. Not after the next release. Now. Because anyone who wants to hack you no longer has to spend years learning how. They can just ask a model. And yes, use those same agents to defend yourself. The same model that can break in can also secure your systems. The question is which side you use it on. Defense is cheaper now than at any point in history. But so is attack. Whoever understands that first will have the advantage.

Bill Gates is right: unpleasant surprises lie ahead. But he’s also right about something else he talks about less often: for the first time, we have tools that can anticipate these surprises—if we use them properly and quickly.

We are the first generation in human history that can see a sunset on another planet. Not in a movie, not in a rendering—in an actual image transmitted from another planet. Keep that in mind when you read about 8.5%, about two hundred dollars for the impossible, about the mathematicians’ letter. The same curve that carried us to other planets is now shooting straight up. It isn’t growing. It’s surging.

We are the first generation that will see a sunset on another planet. We are also the first generation that might not make it through next year if we keep thinking things will somehow sort themselves out. They won’t. But that doesn’t mean it has to end in catastrophe. There’s a window. It’s narrow. It’s open right now. If you have IT tools and critical infrastructure, this is your responsibility. Not tomorrow. Today.

Yan Krivonosov