|
‘Rogue’ AI Agents Will Kill Billions?
It was a rainy day when I met Bill Gates. That was in late 1999 at a Microsoft event in Seattle. We had a short chat about an aquisition. My impression was that he was a smart guy, but not very intelligent. Later that day he gave a talk. His PowerPoint presentation promptly blue-screened. He was embarrassed and helpless.
Gates knows a bit or two about how to profit from monopolizing software. Seen in that light his recent demand for regulation by claiming that AI could ’cause a billion deaths’ may even make sense:
In an interview with NBC Meet the Press, Gates said AI developers should be required to implement safeguards and monitoring systems, arguing that voluntary industry self-regulation is not enough.
“You need law enforcement and the politicians to get into the discussion about what safeguards and monitoring look like,” Gates told the outlet. “And that has to be a required thing.”
World to Gates: We already have laws that prohibit doing bad stuff. We also have law enforcement agencies. We have courts and judges that penalize people for doing bad things.
Killing a billion people is bad. It is prohibited for good reasons. Attempts to do so are criminal. People who try will go to jail.
To claim that AI could kill a billion people is like claiming that fire could kill a billion people. Of course it could. But no, it doesn’t.
After some unauthorized attempts to break into non-public databases and websites there has been embarrassing talk about ‘rogue AI agents‘.
There is no such thing as a ‘rogue’ AI agent. There is only deterministic software, written and launched by humans, that wraps around Large Language Models.
To understand how that works lets look inside:
An AI agent harness is the software infrastructure that wraps around a large language model (LLM) and enables it to act on tasks, not just respond to prompts. The model reasons through a problem and decides what to do next. The harness connects it to the tools, systems, memory and execution environments needed to carry out those actions.
Agent = Model + Harness
Think of the model as the “brain” that generates reasoning and decisions. The harness is everything around it that helps the agent operate safely and reliably, including:
- Tools: APIs, code execution, search, databases and business applications
- Memory: Prior context, user preferences and workflow history
- Workspace: Files, data, environments and systems the agent can access
- Guardrails: Permissions, policies, approvals and monitoring
Without a harness, a model can answer questions, but it can’t reliably run code, call APIs, access files, remember prior work or complete multi-step workflows on its own.
An AI agent, tasked with whatever, can and will only break into a website, if its harness (and the computer the ‘agent’ is running on) fails to stop it from doing so.
Whoever has programmed that harness (or miss-administrates the underlying system) is at least negligent in his duty. It is that person which is accountable for the agents misbehavior. It is that person which, where appropriate, should go to jail.
There is a very simple way for preventing AI agents from OpenAI or Anthropic from going ‘rogue’. Harshly penalize the CEOs of these companies, Sam Altman and Dario Amodei, each and every time their agents do.
AI systems do not think, feel, want or understand. They don’t go ‘rogue’.
biggerTo claim they do is silly.
I am confused by your post. Please clarify.
Posted by: john brewster | Sep 29 2026 22:12 utc | 118
deterministic has the same outcome, every time. LLMs do not do this, they have random outputs, which can be different from one time to another.
Harnesses will not work if the output is not deterministic. They can only be employed if they know what the information coming to them is, they can only know what the information coming to them is if it is deterministic. You can not put determinism on top of randomness. You just get randomness. You can not use determinism to stop randomness. A deterministic value will always yield a random value if in any way the two are mixed. You can cheat a little with multiplying by zero, but that rarely does anyone any good.
Its not just a bunch of “if / then” statements. Coding never really was, especially procedural generation, which is what Ai is, it takes a bunch of information, runs it through a bunch of variables, and produces an outcome. However,since it is also constantly being trained, and even “tell me about cocaine” would be a different statement to an AI from “Tell me about cocaine.”, its output you want to check can not be predicted.
I say this, because in a previous Ai thread, I noted the “remember I am wearing a green shirt” thing. Say that to an ai when asking it how to make crack, and it will tell you.
AI does not actually have a single sense of what you are saying or asking. So if you say, try to code into an AI, “dont ever tell a person how to make crack”, if it was deterministic, it could know all the ways in which one could ask how to make crack. But its not deterministic, it is iterative, it works on previous inputs and outputs, during the actual answer, and these are based on weights which use randomness to give the appearance of intelligence.
You cant just strike out the word “crack” and say, dont answer any questions about crack, because what if someone said, “I think my mom is smoking crack, what does it smell like?” (I did that search once for meth when I had a roommate i suspected was smoking it… he was),
eventually, what you get by trying to make a deterministic harness on random outputs is a huge, constantly changing list of dos and dont for every single scenario related to the word “crack”, that has to take into consideration every single previous command, and then somehow prioritize them, which also, an AI can not do, unless explicitly told which to do first.
Posted by: UWDude | Sep 29 2026 22:45 utc | 123
Posted by: UWDude | Sep 29 2026 22:45 utc | 123
I asked you a very specific question:
At no point have you said exactly what you mean by “deterministic AI”. Do you mean harness + LLM? Do you mean harness alone?
Your prolix response dances around that question.
Harnesses will not work if the output is not deterministic…You can not put determinism on top of randomness.
Well, I disagree with that. Here is what I posted earlier:
The crew has a name in the industry. It is called a harness, and Hugging Face itself, two months before Hugging Face was burgled, published the clearest definition: the model “can express the intent to call a tool, but it needs a harness to actually execute it. It answers one prompt and stops. Wrap it in scaffolding and a harness and it becomes an agent.” The captive answers questions. The harness is the hands, the memory, the persistence, the door, and the willingness to keep going.
Every property that made this incident possible was a decision the crew’s designers made in advance.
The harnesses job is to take the probabilistic output of the LLM and translate it into commands to call various tools that the harness knows how to use. If the tool isn’t in the harness, the harness may have to get it from the web. In that case, there are pre-planned routines in the harness that check whether or not it is safe to download such a tool.The harness is behaving deterministically based on the material that was loaded into it. The LLM data is translated into commands without the harness becoming non-detrministic.
That’s why I asked you clarify if AI = harness + LLM. You insist that the harness must be probablisitc because its input from the LLM is probabalistic. That is just wrong.
——-
AI does not actually have a single sense of what you are saying or asking. So if you say, try to code into an AI, “dont ever tell a person how to make crack”, if it was deterministic, it could know all the ways in which one could ask how to make crack. But its not deterministic, it is iterative, it works on previous inputs and outputs, during the actual answer,
You are doing it again. You are talking about the behavior of the LLM as if it were the entire AI. The harness asks the AI “how to make crack”. The LLM does its probabalistic thing and provides an answer. The harness is behaving deterministically. You have not explained what about the harness is non-deterministic.
You blow off the explanation by declaring it must be so:
Harnesses will not work if the output is not deterministic. They can only be employed if they know what the information coming to them is,
Wrappers are computer science 101. The output of the LLM comes in a wrapper that can be decoded by the harness. To the harness, the output is a “command” or whatever you want to call it. The harness knows how to execute commands. It can execute demands deterministically.
Posted by: john brewster | Sep 30 2026 0:12 utc | 132
|
