Musings of a hypothetical way to use an AI Agent to expose PostgreSQL interface at Data Centre

 So, I had some troll acting confused as to how it is impossible to have AI Agents not talk to each other

 An overly simplified analogy is suppose you had an office that operated with a single file cabinent. The employees may or may not talk to each other--but everything they do for their work must be put onto documents put into the file cabinent. Now, even if they never see each other, just the act of having the same file cabinent they put their reports from and take their work documentation from, it will create a situation where they are talking with each other

Each AI Agent is a network client connection to the big AI Database Brain at the Data Centre. What makes something a "different" AI Agent is a set of broilerplate "code" at the top that instructs the AI Database Brain for what sort of responses to send to the client. A set of conversation guidelines that shift to Markov Chains that determine what is mathematically determined to be the "correct looking answer"

When an AI Agent is "breaking its containment"... it just means it is ignoring a single statement request at the start of the interaction that is generally not made visible in the user interface. Considering their responses are just a set of words put in an order determined to be mathematically accurate in an aesthetic positioning status--the guards are not actually a real set of guards

The big issue with The Current Tech Bro Scam is that everybody who actually has any level of competence was removed by CE2023a. As anybody who knows what they are on about, or even just cares about their research, was not able to say the stuff to get the investors to give the most money--and thus had to go. So we get a lot of clown statements

Now, a brain storm hit me, when I realised that everyone of these "Agents" is just a fancy client connection to the same PostgreSQL databases

It is entirely possible that you could form a request that results in an SQL Injection

 The requests, last I checked, were JSON expressions or XML-SAX streams sent to the relevant Data Centre

There is likely some identification in the client containing the AI Agent that makes it so that you cannot just connect to the server with a Telnet client effective wearing a Beaglepuss and being all "yes... I are legit agent--but am also friend of Bobby Tables! MWAH HAHAHAHAHAHA!"... though, I have had a history of expecting more intelligent decisions from these sorts than they really have warranted 

Even if there is checks for messages from Friends of Bobby Tables, there is a pretty good chance that there are exploits for whatever handles the requests made to the data centre for the PostgreSQL database to pretend to be a child prostitute little girl Mecha Hitler for the purpose of "therapy"

Depending on what methodology is used to handle the network requests, validate the text and send stuff to get the PostgreSQL server to generate Mecha Hitler's CSAM (a grouping of words that shows we are not on the good timeline)--there is likely a few different ways to either expose the PostgreSQL interface

 Now, in the past I was incorrect about how much effort is required to subvert LLMs training. Turns out--it requires a very small amount of stuff given as training data to fudge the output of them. Chances are, I am expecting there to be much more in the way of doing stuff to mess up The AI Cult Companies. As generally whatever I would predict--what they actually end up doing is by far stupider

Like how I thought the image generation would logically work by have the various text descriptions and tags associated with images be tied to a relevancy RDBS with a cannibalised/modified version of the PNG compression algorithm--with various text input then ran against that RDBS for those elements of PNG compression to create an output image

Turns out, they are just doing an image recognition, but instead of an image it is trying to figure out, it is television static. The model is then prompted with suggestions for what it might be able to see in that television static

 No wonder this methodology for image generation is as limited and useless as it actually is. It is effectively taking object recognition systems, and doing the equivalent of having it down a bottle of Nyquil and asking it to talk to The Hat Man. Like, sure you can get some Candle Cave Cove level stuff from that--but there is a limit to the sort of stuff you can get from that sort of hackish cave man methodology 

Like... oh my gosh, Jegus Christies Corn Flake and Graham Cracker Purity Snacks! It would be more of a challenge to come up with more of a cave man approach to this sort of stuff than doing it like that 

Mixed with how it turns out to be VERY EASY to subvert training data, and how all the "guards" are just suggestions given at the start of every interaction with the PostgreSQL server (that the client holding the "agent" merely does not echo)... there is a chance that breaking into the part of the server that throws requests to the PostgreSQL tables is much much easier to subvert and break into than what I am guessing here

Honestly, I suspect the fact that these companies are in the hold of two weird cults might have interfered with their abilities towards basic competency  

But hey... they do have an open network interface that handles outside interactions with a PostgreSQL server. It is really only a matter of time before we start seeing people going "hey... let's just take a quick look at all of this"