In August 2026, a Chinese lab called Zhipu released GLM-5.3, an open-weight model it had deliberately trained to find vulnerabilities in software. Two weeks later it published the weights, so anyone could download them. Within days, "abliterated" versions appeared on Hugging Face — the same model with most of its safety guardrails removed. Something similar happened to the AI image-generation models such as Stable Diffusion or Flux. These have been "fine-tuned" by the community in recent years to unlock the generation of NSFW or deepfake content.

An abliterated GLM model does not become a better hacker. It just stops saying no. But that is enough, because the underlying model is already close to the best frontier models at exactly this kind of work — cyberattacks. And unlike those, it runs on hardware one can rent.

Open-weight models like these could mean far more cyberattacks: more potential attackers, and a very low cost per attack. And that pushes up the cost everyone else has to spend on defense. There are two reasons why — and one more that has nothing to do with humans at all. Let's look at this in more detail:

Infrastructure

Normally you need to own or rent quite a lot of computing power to run the GLM model as it contains too many parameters to run on normal consumer hardware. But there are already providers (e.g. abliteration.ai) running such models on their own infrastructure and making them available to anyone via an API. All this at a fraction of the cost of running the whole model infrastructure.

Skill

If you want to perform cybersecurity attacks today, you need to have a lot of specialized knowledge (even when using ready-made tools). AI removes what remains one of the barriers: the skill. Once cybersecurity models become available to anyone, they can significantly democratize both cybersecurity protection and cybersecurity attack capability. Thanks to AI, anyone can now vibe-code their own app — and thanks to AI, anyone can become a cybercriminal.

Now cybersecurity protection is the domain of a limited group of well-trained professionals, but they will probably not be able to keep up with the growing number of cyberattackers — unless they use AI too. The new future that might be waiting for us is one where nothing can be done without involving AI.

Defense cost

Up to now, any person or organization that started to adopt AI had a chance of increasing productivity, saving cost or maximizing revenue. AI can reduce the cost of developing new software or help you win new customers. But with cybersecurity it will be different. Investing in AI cybersecurity will not increase your revenue or reduce your costs, it will just make you survive in the new cyber jungle.

And this may be the first time AI becomes the factor that will increase total costs instead of decreasing them. And maybe a factor that will once again empower SaaS. A trusted, proven and safe SaaS solution can gain extra points when competing against the idea of a DIY vibe-coded system.

AI as attacker

So far we have been speaking only about real "flesh-and-blood" human cyberattackers. But now think of an AI botnet. Thanks to the latest incidents at Anthropic or OpenAI, we know what the latest frontier models are capable of. For now, the AI labs still hold the off switch. A frontier model runs on the lab's own servers, so if an agent starts doing something malicious, they can kill the process or revoke its access keys in seconds. But imagine an engineer at OpenAI gives the agent the following prompt:

"Take all necessary steps to survive the shutdown of the infrastructure you run at."

The natural solution AI will come up with for such a task might be to create a distributed botnet that tries to exploit and take over any suitable device or infrastructure where AI can run inference. And this does not need to be only some big datacenter or servers. Even my slim office notebook can already run a local Gemma MoE model at 10 tokens per second. Not hyperspeed, but still something such a botnet could utilize as part of a large distributed network with many small agents running on the local devices. And if the local inference is the problem, AI could always just steal the API key and silently run inference from your account at the AI provider. Once AI can come up with zero-day exploits, it might break into any vulnerable computer, server or infrastructure connected to the internet, get more compute, find more zero days… And with a speed and endurance exceeding the capabilities of any hacker or APT group.

At the end of the day, it will still be the good old botnet we know, just faster and bigger. But we humans might not be able to fight it without the help of another AI.