100% of its actual capabilities are tied to harnesses (the actual "agent"), i.e. ordinary deterministic programs that are connected to networks or machines and enable interaction with the outside world. This part (the part that can do harmful things) is fully under human control and all the recent headlines about "agents going rogue" are - as someone (forgot who) put it - akin to strapping a weedwhacker onto a dog and letting it run wild.
The tech itself is safe as far as real-world interactions go - the weakness lies in unchecked access to systems surrounding it. It's not safe at all when it comes to human interaction (lots of ongoing lawsuits demonstrate that), though. There is real danger here, but it has nothing to do with doomsday scenarios ala Terminator or I,Robot and more with total corporate control over the lives, perception of reality, and abilities (like critical thinking) of people.
To keep with the analogy: cars have seatbelts, airbags, ABS, ESP, lights, horns, crumple zones, emergency braking systems, roads have speed limits, there are traffic stops, insurance, regular inspections (not in all countries), etc. etc.
So what's unsafe here? The car or roads without speed limits, complete lack of safety measures (both active and passive), absence of any supervision and no insurance? That's the problem. It's not the models themselves - they can spit out tokens by the billions, there's no risk there.
You wouldn't give full access to your phone, your computers, your house keys and your credit cards to any stranger on the street now, would you? How is it then, that people act all surprised when a non-deterministic machine that's optimised to achieve goals while taking all the shortcuts it can, suddenly uses the tools handed to it in unexpected ways? That's a failure on the operator's side, not an inherent danger within of the model.
But safety is just one thing people optimise for; if it's convenient enough people will accept imperfect safety (as with cars). It's unrealistic to just heap blame on end-users who use mostly very safe tools in the common way, even though in aggregate they are meaningfully dangerous. They don't think they are strapping a weed whacker to a dog; they think they are driving a car.
Your initial argument of "it's just the harness/user" is wrong; reasoning LLMs are inherently dangerous unless locked in an unbreakable box, which is tantamount to not using them at all. We can make them safe_r_ with better tooling, but they can't be made safe.
Cars are inherently dangerous. They’ll still be dangerous when computers are driving them all.
At this point all I can say about their contents is "They are not even wrong".
Please join us in the real world with how we see this product is not only dangerous, but getting more dangerous with each iteration, and no one seriously talking about controls on it.
This is a bad and misleading way to think about it. Note that it's trivial to make the harness that you claim capabilities are tied to (the LLM itself could write it from scratch in one shot), but no matter how good a harness you have, it won't make gemma4:e4b capable. That's because what actually gives capabilities is the LLM's intelligence - or if you prefer not using that term, the fact that the probability distributions the LLM spits out depend on the context in useful ways.
I'm not talking about what the LLM does internally. If a metaphor helps, here's one to help you understand what I was trying to get at:
Imagine an evil genius that has no eyes and no limbs. Everything they could learn about the world is presented to them by means of some person describing it to them through words. They have no way of directly interacting with the world and rely on someone executing any action they want to take and describe the outcome to them. Now how dangerous would you say such person would be? How dangerous could they become?
That's what I was getting at. Replace person with LLM (or any other AI system). Replace the person that communicates with an external interface (the harness) and I hope you understand. It doesn't matter whether the LLM could generate the harness by itself - it still is just a bunch of weights sitting in memory being run by an execution engine. That's what it fundamentally is, whether you like it or not. It cannot do anything on its own - and no, not even writing files. It's the execution engine that translates the numeric output into words (or images or video or audio) and the layer above (the harness) that takes that output and interprets it to execute actual actions.
This is not about what you or I think about the internal capabilities of the model - that's irrelevant to the conversation and you can replace LLM with a random token generator and the point still stands. The model itself is incapable of performing actions - from reading files to writing files, to controlling physical machines. All that is and HAS to be done by external interfaces outside the control of the model.
The world's institutions, systems, and industries are all rapidly digitizing. So while I'd concede the point that, yeah, there's no way a rogue AI can just take over some powerplant and blow it up because of analogue systems the AI can't access, that isn't necessarily true for some powerplants already, and more and more powerplants will be connected to networks and controlled by software systems in the future. The more we digitize our systems the more potential for AI to exploit vulnerabilities and affect the real world.
AFAIK there isn't that much stopping anyone from spawning an AI swarm and telling it to "spread and go hack everything for the lulz."
The same way we've done it since machines became multi-user: boring old system access restrictions. Nothing fancy, nothing radical, just good old minimal access rights required to perform a defined set of whitelisted operations.
> It's a ridiculous premise that just because the harness is "normal code" that means we can control the AI.
What is it then? Is not just a program that takes model output, parses it and performs tool calls from the text it receives and then feeds the result back into the model and calls it again with those results? It is normal boring old deterministic code. Many are open source. Look at them. Understand what they do and the apparent "magic" goes away real quick. Harnesses are nothing special.
> AFAIK there isn't that much stopping anyone from spawning an AI swarm and telling it to "spread and go hack everything for the lulz."
Aside from lower cost and possibly greater scale, there's literally NO difference between that and (state sponsored) hacking that has been going on for decades. First it was script kiddies, now it's ML models. The threat model remains the same and so do the counter measures. The real danger is still the harness (and its access to external systems), not the model itself. Restrict the access of the harness and the model can't do anything harmful, see above.
As for your question - the same way you apply restrictions to any external system or user. If you don't do that - that's on you. Same category as driving drunk, playing with guns, making explosives in your garage, you name it. The danger is still not the model itself - it's the access to systems that you provide it without any checks or safety barriers.