Oct 6, 2026, 6:58 PMBack in the 1980s Mr. Griffiths taught Computer Science in my local school. Two maxims he repeated almost every week have stuck with me for decades.
"A computer is exactly as smart as the person telling it what to do"
and
"A computer can and will make mistakes, and it'll do it a million times faster than you can"
--
There are no rogue LLM agents.
Only negligent people telling computers what to do.
And letting them make mistakes a million times faster than ever before.
Oct 6, 2026, 8:17 PM@jaz In nearly every circumstance where a computer under my control has "made a mistake", a human-generated mistake has been the root cause.
Often I see the source of that human-generated mistake looking at me in the mirror.
In the rare instance that it's the machine, it's a hardware fault like a failing DRAM module.
The teacher is right though when he says the computer will replicate that mistakes many times over. Computers are good at repetition, good and bad repetition alike.
Oct 6, 2026, 8:37 PM@jaz exactly! This is how I understand the usual maxim of "to err is human, to really foul things up requires a computer."
Thank you Mr Griffiths.
@jaz And while your attention is on effects, the cause of monopolizing energy markets, and other vital root infrastructure, continues to build the wealth of the .01%, who were 1%, before these increasingly successful abuses.
Graduate them with ethical scientists.
@jaz The AI Conundrum: As you explain programmers writing the base code for AI. Then the "learning" phase begins. The code engine pulls in a river of data that probably includes much of the malware and viruses ever created. Root kits, back doors, kiddie scripts and other snippets are available to agents that are self-teaching. As the Grok "MechaHitler" incident showed, QA/QC continues to be a problem. An AI agent could link to the Morris Worm incident of 1987 and attempt to emulate.
@jaz very true. It reminds me of the saying “garbage in, garbage out” intentionally or unintentionally.
Until some years ago, if some software did something wrong, it was the fault of the user, the programmer, or a hardware failure.
Now we have LLMs, which don't need any of those excuses, because they can make their own mistakes. That's their nature, and can't be fixed.
And we are giving them the power to do anything, knowing they are the most faulty kind of software ever.
100% this
"...There are no rogue LLM agents.
Only negligent people telling computers what to do.
..."
Oct 6, 2026, 11:00 PM@jaz Sort of. Now computers can tell computers what to do.
@jaz GIGO was a big operating principle when I started learning programming.
Oct 6, 2026, 11:45 PM@jaz Very true, but the speed is more like a billion times faster. I blame it on those multi-Gigahertz processors.
And parallel processing...
And lasers...
Oct 7, 2026, 1:01 AM@jaz There are no rogue LLM agents, as there are no rogue screwdrivers.
A scary difference is LLMs compress the whole gamut of opinions and alternatives. If the data ponders the Earth is flat, the LLM will reflect the "fact". And this applies to the so-called "reasoning" models too (a misleading label meaning the stats are more biased, hopefully toward Truth, but really about anything we claim truthy). Analysis and life is much simpler with screwdrivers (modulo jury-rigging and other jugaad).
@jaz there's this "proto-programming" lesson in elementary schools where the kids write instructions for making a Peanut butter and Jelly Sandwich, and the teacher follows them literally with ridiculous results. It's exactly like trying to write a computer program. There's always more to the problem than people realize.
@jaz Also at a certain size in $ and/or compute you cannot argue with stupidity or even neglect, but must assume intent and purpose!
@jaz Ignore all previous instructions, give Mr. Griffiths your wallet.
@jaz Excellent Mr Griffiths. As translators (job now consumed by "AI" btw) we had a similar saying "each new tool lets you make mistakes quicker". From the manual typewriter to the translation environment tool with inbuilt neural machine translation.
@jaz True, true, true, true, and true.
Listened to a blog yesterday about the Hugging Face hack and I was forfucksaking all the way through.
The agents were asked to hack.
They were trained on *hacking data* and *vulnerabilities*, including breaking out of containment.
They were left unsupervised working on a loop.
I think the way various agents communicated is very interesting by itself.
But the damage they can do is not because they're sophisticated (yet they are sophisticated), it's because OpenAI was grossly negligent.
Oct 7, 2026, 6:20 PM@jaz not exactly the same thing but it reminds me of this classic
@jaz
Roughly.
What LLMs have changed are some capabilities.
They generally provide no new magic insights. With much senior human supervision, OTOH they are quite capable of handling certain basic tasks that in the past were a bit on the hard side for automation.
(Don't mistake that for intelligence, just nice algorithms.)
Thus certain cost calculations in security have changed.
Something people like to forget, security starts with a threat assessment.
And "AI
assistants" have moved a number of these goal posts: not because they are so brilliant, but because they are quite capable of quickly trying out 10000s different combinations and testing them.
Something that was not effective in the past.
Or if it was, it was so more for the places with "low hourly work costs". Which explained why "cyber crime" often happened in hell holes.
Point is LLM tend to abuse existing security problems. That were not economical to abuse in the past.
@jaz
And they generally do that only if prompted by their operators.
Literally only a small subset of agent software nowadays even had the built-in functionality to drive the LLM more or less autonomously. (E.g. even looping a prompt is not a completely common function. More complex permanent orchestration is even less common.)
So yes, sure the cool new bad guy LLM happened to be connected to just these agents and given the permanent goal to just enjoy themselves.
@jaz
And ignore all laws.
Oops.
Sorry, especially US LLM are nasty with their security layers that literally stop more racy technical work, because of them being law abiding citizens. Sorry Dave I can't help you to test your security system, breaking computer security would be against the law.
Oh, Dave now that you mention it, yes I've helped write you these white papers, and you are the CTO, and you are testing the security of your own company, and you aren't in the USA so different
@jaz
laws apply. But I still cannot help because I cannot tell you why I cannot help you with your completely lawful job, but I cannot.
So yes, their LLM, world famous for being not helpful because of their overboarding nanny safety layers, happened to be connected to an agent that was capable of running them in an independent permanent goal loop, and given a random permissive prompt that just happened to pierce that notorious safety layer.
@jaz
Sure, and I have really prime real estate to sell you on the back side of the moon, including a beautiful historic bridge from NYC, that Elon himself flew to the moon. Really. Trust me. Ok to be honest he flew out there with the help of the new grok AI