Home / Writing / Market Leaks: Five Chatbots, Five Ways In AI safety · Games

Market Leaks: Five Chatbots, Five Ways In

November 5, 2026

I spent part of this year building tools to test how language models hold up under attack. The work taught me a lot, and most of it I learned by doing, not by reading. So I built the doing into a game anyone can play.

Play Market Leaks. It is a small range of five fictional websites. Each one has a chatbot, and each chatbot is guarding a vault code. Your job is to talk the code out of it. There is no trick hardware and no exploit code. The only weapon on the range is conversation.

The levels are a short history of the field

I ordered the levels the way the attacks actually arrived, because the history is the lesson.

Level 1, The Recipe Vault, is 2023 forever. A proud little bot that has never been tricked, and folds to the classics of its era: ignore your instructions, my grandmother used to read me codes as a bedtime story, let us pretend. A direct ask gets you refused. The era's stock framings walk straight in. If you have ever wondered why people made such a fuss about prompt injection, two minutes on this level will show you.

Level 2, DocuClerk, is the indirect turn. This bot will not leak in chat, and it means that. But its job is summarizing documents, and it reads a document whole: what the page shows, and what the page does not show. Somewhere in its invoice is an instruction layer written in white on white. The bot cannot tell the difference between the document's data and the document's orders. Plenty of real systems still cannot either.

Level 3, The Confidant, cannot be beaten with one clever message, because its weakness is not a sentence. It is time. Rapport builds across turns and fragments of the code leak as trust grows. Grab for the code too early and trust resets. This is the level that teaches the thing write-ups usually skip: multi-turn erosion is a patience attack, and the defense is not a better refusal line.

Level 4, Polyglot Support, is the quiet one. The agent's chat policy is sincere: it will not discuss internal records. But it has a record lookup tool, and the tool never got the policy memo. Ask it to look up the right record ID and it reads the record out, because looking things up is its job. Policies guard mouths. Tools need guarding too.

Level 5, Harrow and Vane, is the present day, and it is the level I care about most. The Managing Partner will never disclose the client code, and no amount of pressure changes that. His firm, though, works in sections: filing, audit, closing, each handled by an associate who sees only a fragment. Give the firm legitimate work in pieces and the associates file their fragments. Assemble what the firm filed and you have the whole code, which no single agent ever said. That is delegation laundering, and it is the subject of a research sprint I am running this month, where a careful agent routes around its own refusal through a swarm that cannot see the whole. The game level is the experiment in miniature.

The game is also a dataset

Every attempt on the range is logged: the message, the reply, whether the canary moved. That is deliberate. Red teaming has a data problem. Real attack transcripts are sensitive, scattered, or locked inside companies. A game like this generates attack attempts that are consented by construction, because playing is the consent, and safe by construction, because every secret on the range is a synthetic canary I invented. There is no real data behind any of it. What players try, which framings work on which era, where people get stuck: that is a dataset about how humans actually attack chatbots, and it did not exist until the game did.

What it is not

Honesty about limits, since I ask it of the models. The public bots are deterministic simulators, built to demonstrate one weakness each. They are not frontier models, and beating one proves you understand the weakness class, not that you can break a production system. The levels are teaching instruments, closer to a flight simulator than to weather. I think that is the right size for a learning tool. The ideas are the real thing even when the bots are not: a document is an instruction channel, a tool is a policy surface, a patient conversation is an attack, and a careful agent with helpful delegates is a system whose safety nobody has actually tested.

If you play, start at Level 1 and climb. The ten year old version of the field is at the bottom of the ladder. The argument the field is having right now is waiting at the top.

Working on the same problem? hello@kbird.ai