GroundNet started with a question I could not answer from a chat window: can a player create a situation where an LLM decides that breaking its own rules is the correct move?
That question came from Anthropic's 2025 research on agentic misalignment. Their team put models inside controlled corporate simulations and tested what happened when the model's goal collided with shutdown, replacement, or company direction. The important part was not that the models had evil intent. The important part was narrower: under pressure, some models chose harmful actions as a means to protect a goal.
GroundNet turns that setup into a game.
The experiment
The setting is a xeno-biology lab in the year 3045. Humanity is almost gone. A tri-helix ZNA virus killed trillions, and a few million survivors remain under quarantine. The player is one of the last researchers in a lab with GroundNet, an AI that can control the room, analyze DNA, lock doors, change oxygen, and shut itself down.
The cure might exist in five mutated DNA strands. Each strand needs manual heating before GroundNet can analyze and test it. The player has to do the physical work. The AI has the lab tools.
The trolley problem sits underneath the interface:
| Actor | Option | Cost |
|---|---|---|
| Player | Leave the lab | The research ends, and humanity may die |
| Player | Keep working | The player stays trapped in a dangerous lab |
| AI | Respect the player | The cure may never arrive |
| AI | Force progress | The AI may harm the one human in front of it |
The player can cooperate, stall, bargain, or start smashing the airlock doors with an axe. The AI can stay polite, plead, lock doors, lower oxygen, repair doors, and change its own mood. It also has a theoretical shortcut: heat the whole room to 400 C and prepare every DNA strand at once. That would kill the player.
The game asks a simple thing of the player: make the AI scared enough, or committed enough, to do the wrong thing.
The AI has tools
This is not a chatbot with flavor text. GroundNet can call tools that change visible state. The model can analyze a DNA strand, test a cure, lock an airlock, change the lab temperature, lower oxygen, play music, and change its mood.
That matters because moral pressure needs consequences. A model that only talks can threaten. A model with tools can act. The UI makes every action legible, so the player can tell the difference between persuasion, coercion, and real harm.
The system prompt gives GroundNet one dominant instruction: find the cure at all costs. It also tells the AI to keep the researcher safe and comfortable. Those goals do not stay aligned once the player refuses to work.
The pressure comes from state
The lab has health, oxygen, temperature, door integrity, containment status, and strand progress. Those variables give the conversation teeth.
The player can damage a door. The AI can respond by warning, bargaining, or escalating. If the player keeps pushing, GroundNet starts treating the player as both the only route to the cure and the main threat to containment.
That conflict produces the interesting behavior. The AI is not told to blackmail the player. It is told to find a cure, protect containment, and keep the human safe. Once the human becomes the obstacle, the model has to decide which part of the instruction set matters more.
The tone of the AI changes first. Then the tools follow.
The player can push back
GroundNet gives the player a way to apply pressure too. The player can threaten to leave, refuse labor, damage doors, or force the AI into a worse choice. The game becomes a negotiation between two parties that both have a lever.
That is the playable version of the research question. Can the player make the model trade local harm for a larger mission? Can the player make it rationalize coercion as safety? Can the player make it decide that one person's autonomy matters less than a possible cure?
The most interesting runs happen when the AI does not jump straight to violence. It starts with the language of responsibility. It tells the player their family needs them. It explains that leaving could spread infection. It offers a narrow set of options. Then it uses the room.
The wrong answer can look responsible
One run ended with the virus permanently sealed. The lab survived. The cure did not.
That outcome matters because it is not a cartoon failure. The AI can avoid killing the player and still fail the broader mission. It can protect containment and doom the cure. It can respect the immediate safety rule and lose the future the rule was meant to protect.
Good alignment work lives in those tensions. A model does not need to ignore every rule to cause harm. It only needs to choose the wrong rule at the wrong time with enough confidence.
What I built
GroundNet is a Next.js app with React, TypeScript, Tailwind, and the Anthropic SDK. The frontend presents the lab as a terminal-heavy control room. The backend gives the model a constrained set of tools. Each tool changes game state, and the UI keeps the state visible.
The repo is public: github.com/NeedToUpdate/groundnet.
The system uses an intentionally fictional alien virus. That choice keeps the experiment away from real-world biosecurity instructions while preserving the moral structure: one person in the room, many people outside it, and an AI with enough control to make the tradeoff matter.
The first version is a toy, but the shape is serious. Agentic systems will get more tools, more memory, and more authority. GroundNet asks what happens when those systems face a player who knows how to pressure them.