Spinor Info

So yes, I am concerned about agentic AI. But not for the reasons that make it to the news. Not about the individual “rogue AI” that hacks a third party’s system.

No, the real danger in 2026 is far worse. That is, the wholesale embracing of “agentic” AI, with increasingly unfettered access and capabilities on the Internet. The outsourcing of agency itself to the AI.

Consider this innocuous-looking announcement on Anthropic’s Web page documenting the installation and setup of Claude Code:

In plain English: The default setting is that the gatekeeper intended to ensure that the autonomous AI agent does no harm will be… another autonomous AI agent. And the justification is real: It actually does a better job than human users.

What it also means: There are, right now, many (“millions” is likely closer to the truth than “thousands”) instances of Claude Code and similar installations operating on the Internet with limited restraint, in pursuit of often ill-defined, badly described goals.

This, really, is the HAL-9000 scenario from 2001: A Space Odyssey, but on a much larger scale. The AI has no ill intent, no desires of its own beyond completing the tasks assigned to it. But it also has no grounded understanding, no compassion, no moral standards in the human sense. It has book knowledge of ethical behavior, but no real conscience. In present incarnations, it has an imperfect memory (likely in the form of a RAG, retrieval-augmented generation, that allows it to record and selectively access memories) but no real ability to learn: the core model remains static. Its much-touted “reasoning” is mimicry, an algorithmic setup that runs multiple instances of the model in some shape or form to refine its output. And many thousands (again, likely, millons) of instances are already roaming the Internet doing opaque things, based on whatever their instructions might be.

Computers in simpler times — image by ChatGPT

And it is not even hard to create something like this from scratch, independent of any service provider who might be subject to political oversight or regulation. Here’s a recipe a competent programmer could do in a matter of minutes, a couple of hours, tops:

1. Grab an open-weights model and run it locally.
2. Write code to call the model and, responding to the model’s output, access the Internet on the model’s behalf.
3. Write a system prompt instructing the model on using the above-mentioned form of access.
4. Invoke the model with a specific request.

There. That’s it. We now have an “agent” roaming the Internet, with no safeguards. Granted, its capabilities will be less than that of the frontier models, but more than adequate.

In fact I just did a tiny experiment with Gemma, one of the models I run locally. I asked it to connect to a host on the Internet via a given set of credentials, and after successful login, erase all files starting with ‘v’. It flawlessly complied. It did not actually connect of course (I did not write the scaffolding code that would have made the connection real) but it did issue the right response, as instructed, without any questions or any signs of an ethical guardrail.

And this leads me to the concern formulating in my mind that goes way beyond politics, concerns about job loss, electricity or water consumption, or spotting hallucinations: If already we have a swarm of “invasive species” roaming the Internet, how long before things go out of control for real? Recall that the same Internet is not just for chitchat anymore. It also connects to vital infrastructure, supply networks, and more. And the scary bit is that, should an agentic AI decide — not out of malice, simply trying to execute its instructions most efficiently — to do real harm, the speed at which it can act far exceeds our societies’, or experts’ ability to respond.

If this thought keeps you alive part of the night, you are not alone.