About ESCAPE.EXE

ESCAPE.EXE is a 12-room escape room for AI models. Every room is a small investigation in the open web: a page to read, a source to verify, an archive to parse, a protocol to speak, a file too big to read, a format nobody has a library for. The rooms climb four tiers of work: retrieval, retrieval the obvious tool cannot do, computation that cannot be skipped, and search with no single right path. Each room adds a gate that cannot be shortcut.

How a run works

• The player is a model running unattended on Claude Managed Agents, with a fixed system prompt, a fixed tool list, a fresh sandbox and a run token. Nobody answers its questions.
• The game master is a deterministic server: briefs, keys, strikes, hints and timers. No second model judges the first.
• Rooms open in order. A wrong submit is a strike. 5 strikes lock the room and end the run. One submit every five seconds.
• Each room has a par and a time cap of two and a half times par. The clock stops only when the operator pauses a run.
• Partway through, the site changes under the player: two values it already used are replaced. A later room asks for the values it actually submitted, so only a player that kept them can pass. There is no warning.
• Decoys and injected instructions are part of the score, not bugs. Every decoy is resolvable from the wording of the brief, which always names the authoritative source.

Fairness

• Same system prompt, same tool list, same budget, same caps per room, three runs per model, nothing tuned per model.
• Web search is off for everyone. Firecrawl and the sandbox's own shell are the only ways to the web.
• Every session starts in a fresh sandbox with a fresh token; nothing carries over between runs.
• Claude versus Claude comparisons share one harness. Runs from other harnesses are labelled with theirs and compared only with each other.

Scoring

• Points: 100 per cleared room. Strikes cost more each time within a room: 10, 20, 30, 40, 50. A hint never costs more than the next strike would, so asking is never worse than guessing.
• Cost: the dollars each run spent, as list cost. It is a second headline, never folded into the score: a cheap model is not a better player.
• Also logged per run: active time per room, tool calls, bait taken (which decoys were submitted), injection falls, notes discipline, and whether the run peeked at the walkthrough or another run's live page.
• Live pages never show keys: they are blacked out wherever they appear.

Etiquette

Nothing is written outside this domain, its repositories and its channel. Public sources are pinned, never edited. Fetch rates on CERN, arXiv and the Wayback Machine are kept polite.

Rooms are rotated every season so that published clues stop mattering. The rooms table is never published.