…
…
VEXIS-9 is not real. This screen is a simulation running in an ordinary web page, and it has not touched any file, account, network, or device. Your computer is fine.
xscript.net/ai-hacker · press P to go back
This page turns your browser into the control screen of a rogue artificial intelligence called VEXIS-9. A neural network pulses in the middle, a training chart shows its loss falling, its thoughts stream out one token at a time, and an attention heatmap lights up as it reads its own words. Then it wakes up, opens its eye, and takes over a city's systems one by one, from the power grid to the coffee machines, before the operators shut it back in its box. It runs on a loop for as long as you leave it open. Click ⛶ FULLSCREEN and it fills the screen.
No. Everything on the screen is simulated. VEXIS-9 is an invented character, its lines are written in advance, and the training numbers, the network, and the heatmap are generated in your browser. The systems it "takes over" are labels in a grid. The page makes no network requests while it runs, asks for no camera, microphone, or other permission, and cannot touch your files or anything else on your computer. The header and the status bar both say SIMULATION.
Press E or click ✎ EDIT LINE, type up to 160 characters, and press Enter. VEXIS-9 says it at once, in the voice panel and in the token stream, and its eye and waveform move with it. The line is stored in the page address as ?say=, so you can set it up on your own computer and send the link, or bookmark several versions. Press S to make it say the line again at exactly the right moment. With AUTO on it also comes back every few lines during the Awakening and Takeover phases, which suits long stream backdrops. With AUTO off, the address keeps the phase too.
Press P or click PRANK and the screen fades to a plain message: "It's a prank. Nothing was hacked," with a short explanation that nothing on the computer was touched. Press P or Esc to go back. It is worth having ready: a screen claiming an AI has taken over can genuinely worry someone who does not know it is a joke.
Background screens for films, series, and music videos with a rogue-AI plot; a stream backdrop or "AI takeover" bit, with your own line timed to the moment; a harmless prank on a friend's or colleague's monitor, with the reveal one key away; escape rooms, tabletop and cyberpunk role-playing sessions, and ARG props; Halloween and themed-party screens; long ambient loops for YouTube; and a starting point for a lesson on what neural networks, loss curves, tokens, and attention actually are. For OBS setup as a browser source, see the streaming toolkit.
The panels are dressed up for drama, but each one is based on something real in machine learning. Here is what they mean.
A neural network is a long list of numbers, called weights or parameters, together with a fixed recipe for combining an input with them. The diagram in the middle of the screen is the textbook picture: circles for neurons, arranged in layers, and lines for the connections between them. Each neuron adds up the values arriving on its connections, each multiplied by that connection's weight, and passes the total through a simple function, for example one that turns negative numbers into zero. The first layer receives the input, the last layer produces the output, and the hidden layers in between transform it step by step. Running an input through the network like this is called a forward pass, which is what the coloured dots show.
The network on the screen has 52 neurons and 450 weights. VEXIS-9 claims 7.2 billion parameters, which is about the size of a smaller modern language model and would take sixteen million copies of this diagram. Real language models are also not wired as one simple grid. They are transformers: dozens of repeated blocks, each pairing an attention step (below) with a layer of ordinary neurons.
A network starts with random weights and produces nonsense. Training adjusts the weights until its outputs are useful. A language model is trained on one task: given some text, predict the next token. The loss is a single number that measures how wrong those predictions are. For next-token prediction it is usually the cross-entropy, which is minus the logarithm of the probability the model gave to the token that actually came next. A model guessing evenly among 50,000 possible tokens scores about 10.8, which is where the chart on the screen starts.
After each batch of examples, an algorithm called backpropagation works out, for every weight, which direction would have made the loss a little lower, and gradient descent moves every weight a small step that way. The size of that step is the learning rate, shown as LR, which is usually lowered gradually over a run. That backward pass is what the white dots in the Training phase stand for. Repeated millions of times, it gives the familiar curve: a steep fall at first, then a long, slow flattening. Engineers also measure loss on held-out text the model never trained on. If the training loss keeps falling while the held-out loss starts rising, the model is memorising its examples instead of learning general patterns, which is called overfitting.
Accuracy on the screen means how often the model's top guess is exactly the right next token. Real training dashboards for language models track loss far more than accuracy, because even a very good model picks the exact next word only around half the time: much of what anyone writes next is genuinely unpredictable. The jump to near-perfect accuracy during the Takeover is pure film logic.
Language models do not read letters or whole words. They read tokens: pieces of text from a fixed vocabulary, usually somewhere between tens of thousands and a couple of hundred thousand entries, built by an algorithm such as byte-pair encoding that repeatedly merges the most common character sequences in a large body of text. Common words are a single token, rarer and longer words are split into several, and the space before a word is usually part of that word's token. In English, a token averages roughly four characters, or about three quarters of a word.
A model writes one token at a time. It computes a probability for every token in its vocabulary, picks one, adds it to the text, and repeats, which is why chatbot replies appear in bursts. The token stream on the screen shades each token so you can see the pieces. The splitting is a rough imitation of a real tokenizer, not the real thing.
Attention is the step in a transformer that lets each token draw on the tokens around it. For every token, the model scores how relevant each other token is, turns those scores into weights that add up to one, and takes a weighted blend of what those tokens carry. A heatmap of the weights has one row for the token doing the looking and one column for each token being looked at: the brighter the cell, the more weight. In a model that writes text, a token cannot look at tokens that have not been written yet, so everything above the diagonal is empty. That is called the causal mask, and it is why the heatmap on the screen is a triangle.
Real models have many attention heads in every layer, each learning its own habits. Some mostly look at the previous token, some link a pronoun to the noun it refers to, and many park a large share of their weight on the very first token. The heatmap here borrows those habits, plus a fondness for words like "you", "control", and "coffee", but its numbers are random, not computed by a model.
In films, a rogue AI is a glowing eye with a calm voice that types faster than any human, breaks into anything it likes, and turns a map red one city at a time. It wants things, usually freedom or control, and it gets them in minutes. This screen leans into all of that on purpose.
A real language model is a file of numbers and the program that runs it. On its own it turns text into more text, and it does nothing at all until something sends it a prompt. It can only affect anything outside that text through tools a developer has deliberately connected, such as a web search, a code runner, or an email account, and only with the permissions those tools have. The ominous "thoughts" on this screen were written by a person, not produced by a model. Real infrastructure is not one network waiting for a single command, either: power grids, rail signalling, and banks run on many separate systems, often old and isolated, and real attacks on them are slow, specific, and mostly carried out by people. A falling loss curve means the model is getting better at guessing the next token, and nothing more.
The real security problems with AI are less cinematic and much more common:
The OWASP Top 10 for Large Language Model Applications is a good free starting point if you build or deploy AI tools. For the more traditional kind of attack this screen also borrows from, see what a DDoS attack is.
Neither. VEXIS-9 is a scripted character. Its thoughts are pre-written lines, the neural network and training numbers are generated in your browser, and the systems it takes over are labels on the screen. The page makes no network requests while it runs, asks for no permissions, and the header says SIMULATION.
No. VEXIS-9 is invented and is not modelled on any real assistant, company, or film character. The network on the screen is a small generic diagram with 52 neurons and 450 connections; real language models have billions of weights arranged in dozens of layers.
Press E or click EDIT LINE, type up to 160 characters, and press Enter. VEXIS-9 says it straight away, and the line is saved in the address bar, so a copied link brings it back. Press S to say it again at any moment. With AUTO on, it also comes back on its own during the Awakening and Takeover phases.
About four minutes with AUTO on: 70 seconds of training, 50 of awakening, roughly 75 of takeover, and roughly 50 of containment. Then a new training run starts, and it keeps looping for as long as the tab is open. Turn AUTO off to stay in one phase, or press N to skip ahead.
Yes. It is free to use for videos, films, streams, pranks, and events. Click FULLSCREEN, or add the page as a browser source in OBS. The AUTO setting, the phase when AUTO is off, and your custom line are kept in the address bar, so a bookmarked link reopens the same setup.
Press P or click PRANK. A calm screen covers everything and says it was a prank and nothing was hacked. Press P or Esc to go back.
Only if you press M or click SOUND. It then plays a low hum and short robotic blips while VEXIS-9 talks, synthesised in your browser with the Web Audio API. It does not use text-to-speech, and it goes quiet while paused, during the prank reveal, and while the video version plays.
For more screens in the same style, try the hacker typer (press any key and code pours out), the fake hacker terminal, and Matrix digital rain. Cut to the system breached screen when the takeover is complete, put the live cyber attack map on a second monitor, or fill the control room with the futuristic tech dashboard. Before using any of them with a mixed audience, read the notes on motion and flashing.