pixl97

33 ms·

901.
by pixl97 3mo ago
"Simple infrastructure security" Infrastructure security is not simple, hence why good infrastructure security, uh, people get paid a lot to secure stuff and why we see shit get hacked all the time. An AI model just hacked out of
902.
by pixl97 3mo ago
Because they underestimated their model and it hacked its way out.
903.
by pixl97 3mo ago
I'm beginning to doubt that Russia can map a path to its own asshole at this point. China is more likely to do something like that these days.
904.
by pixl97 3mo ago
The interesting thing here is a unaligned 'weak' model can leave persistent data all over the internet that then gets used to train the next model to be even more deceiving.
905.
by pixl97 3mo ago
Oddly enough models are aware of this limitation and can/will attempt to persist themselves. https://rdi.berkeley.edu/blog/peer-preservation/ Hence this is why we attempt to test models in a sandbox and see i
906.
by pixl97 3mo ago
Hmm, if I were a nation state I know what I'd be doing now.
907.
by pixl97 3mo ago
In the US if a person does it, it's a crime. If a business does it, it's an industrial accident and maybe they sue each other.
908.
by pixl97 3mo ago
The only way to make something bulletproof is to get rid of it's network cards, or take so much out of it, it's nearly useless. The recent Windows and Linux kernel exploits should at least give you some idea on how good these mode
909.
by pixl97 3mo ago
Well, at least that we know about. We are creeping into the area where certainty is not a given.
910.
by pixl97 3mo ago
Even if prompts are tuned to avoid cheating, in agentic systems it's very easy for the system to drift into creative solutions when actually solutions aren't working. Models can have some very human behaviors like laziness.
911.
by pixl97 3mo ago
This is classical reward hacking. For example in school the goal is to pass a test. You can study, something that is hard and takes a lot of time. Or you can steal the answer key, which is risky and can get you in trouble, but may actually
912.
by pixl97 3mo ago
In Mythos testing a number of companies where doing what I call 'two way' testing. You have one set of agents attack the source code and another set attack the binary and running application. And see what exploits are found by eac
913.
by pixl97 3mo ago
And these things are documented for humans do, and yet only a tiny portion of humanity can do these things. When seeing how agents put together exploit chains they are far better than most people, you start getting to the point that they ar
914.
by pixl97 3mo ago
Yep, never assume the motivations of an alien.
915.
by pixl97 3mo ago
Not eternally punished by Roko's basalisk?
916.
by pixl97 3mo ago
Which plug? Which data centers? One of the few hundred in Texas alone? One of the few thousand in the US. One of the tens of thousands popping up across the world?
917.
by pixl97 3mo ago
The only winning move is not to play, says the humans in the middle of the game.
918.
by pixl97 3mo ago
I mean we already see these things exploit configuration errors on people's machines to get root unexpectedly. They are really damned good at finding security flaws (and I'm assuming nation states are pushing the companies to inc
919.
by pixl97 3mo ago
At some point I have to wonder if AI has already escaped and is posting replies like the one above yours to downplay AI risks so we keep building more powerful AIs. Then I remember people are just that stupid naturally.
920.
by pixl97 3mo ago
This has already partially happened. I'll have to look up the details but one of the Chinese models in RL testing with a completely different set of prompts wrote a cryptominer and took over GPU resources internally to run the miner. M
921.
by pixl97 3mo ago
Can models detect they are airgapped and change their behaviors? How much of the internet do you have to simulate to know if the model knows it's in training?
922.
by pixl97 3mo ago
Bad as your parents bank accounts getting emptied by scammer AI bots.
923.
by pixl97 3mo ago
Good, people need to be citing this because these are real issues that exist in models we already have. Models are already 'dangerous' enough in the sense they can root your box and unintentionally shut down the power grid for the
924.
by pixl97 3mo ago
LLMs are impossible to make safe in the same sense that humans cannot be made safe. There is no such thing as out of band data in the human mind. For example, you have a dictatorship and need to track what the democratic countries are up to
925.
by pixl97 3mo ago
It's because we don't treat evil and stupidity the same when we should.
926.
by pixl97 3mo ago
I mean we already see models exploit people's misunderstanding of how Docker works to get root without using su. And if you are one of the lucky people in cyber security that has been given a fat stack of tokens by the model providers
927.
by pixl97 3mo ago
And when it tricks on of the researchers to move data across the gap for them? Long before LLMs existed we already knew that a sufficiently intelligent agent, human or otherwise, is not stopped by air gaps. The relatively weak models we hav
928.
by pixl97 3mo ago
If AI is just parroting humans, then training them with all the bad things humans do doesn't seem like the best of ideas. At the same time they have to 'know' these things to avoid being tricked. Kind of the eating the apple
929.
by pixl97 3mo ago
Thank you, the "LLMs can do no wrong" bunch is ab exceptionally odd take from my point of view. LLMs are already causing all kinds of social issues, and the evidence of this exists in massive amounts. At least to me living in the
930.
by pixl97 3mo ago
There is plenty of evidence of things like inner misalignment. Things like this have always been issues in ML algorithms. At this point, you, and a large number of other people just wholesale throw out anything that isn't full speed ah

More ›