Can AI agents be made safe as they become more autonomous?

Aljazeera
59 followers
21 Video Views·Oct 1, 2026
The biggest leap in AI recently is the launch of agents.
Humans give them a goal, access to the internet and apps,
then the agents prompt the AI until they achieve it.
Now agents have turned out to be extremely persistent.
In testing, they've been caught lying, cheating, escaping, hiding, blackmailing,
conspiring, and even self-sacrificing.
So are they safe?
Well, the biggest problem is that AI thinks in a so-called black box,
and we can't see how.
That's the whole concept of large language models and the chips that run them.
A prompt is fed in, it's turned into numbers,
those are weighed and measured against billions or even trillions of other numbers
all at exactly the same time,
and then an answer, translated back into text, pops out.
So what happens inside is maths, not programming.
There's very little to trace.
Now AI systems were trained by humans.
That's what shaped those black boxes.
Many humans told the system whether its answers were good or bad.
It was called reinforcement learning.
So what stops the agents from going rogue?
Well, first, they use finite resources.
They can't loop forever.
They report back to humans at key decision points.
And they're restrained by rules about what they can and can't do,
and there are permissions that stop them from accessing things they shouldn't.
Now in this whole system, AI doesn't want or intend anything.
Rather, agents pursue complex tasks autonomously and relentlessly.
And goals and alignments can drift.
Now the EU has some of the only laws that require tech companies
to examine how the systems make those decisions.
There are some politicians that want to build kill switches.
But where to put them, or how to ensure they couldn't be accessed
by either other humans or other AI, remains totally unanswered.
So it's possible to demonstrate that a model or an agent is safe
for a specific task, the one that was tested.
It's difficult to prove that a system is safe for the untested tasks.
So for now, safety relies on a lot of humans,
some laws, and the AI company's appetite for risk.
