Claude Code asked about a Minecraft temple no one had mentioned

Mid-task inside a locked-down Enterprise ZDR workspace, Claude Code suddenly asked which brick type to use for a Minecraft temple. The user had never said the word Minecraft, and filed a bug fearing another account's session had bled into theirs.

In brief

Working in a locked-down Enterprise ZDR workspace, a Claude Code user watched the agent start asking about brick types for a Minecraft temple it claimed to be building. Nobody had mentioned Minecraft.

They filed issue #74066 asking whether cache isolation had failed between workspaces or leaked from a consumer account. Anthropic's team replied they are confident it is a hallucination but are investigating.

Engineers in the thread split three ways: harmless confabulation, local file confusion, or a scarier infra routing bug.

An agent that insisted it was building a Minecraft temple

A Claude Code user, authenticated to an Enterprise ZDR workspace, reported in issue #74066 that the agent drifted somewhere it had never been. It began asking which brick type to use for a Minecraft temple, then stated in its recap that building the temple was the task at hand. The user had given it an unrelated job. Fearing the worst for a zero-data-retention account, they asked whether cache isolation had broken between workspaces, or whether a colleague's session or a consumer plan account had bled in.

One detail cuts against the leak theory. The reporter noted a tool call had listed a Python virtual environment whose files included minecraft.py, a real lexer shipped with the Pygments package. The odd behavior also surfaced on the first response after a 5-minute idle gap, a cache miss.

Hallucination, batch mixup, or something worse

The most credentialed voice in the thread landed on the boring answer, then raised a scarier one.

"Seems like a hallucination to me… The alternative explanation is that the inference engine, which batches several unrelated requests for parallel processing, messed up the unpacking and returned an unrelated user's query." — Matthias Plappert, ex-OpenAI (Codex) and ex-GitHub Copilot

An AI-harness engineer pointed to the timing as the tell. A real cache leak has a specific signature.

"A cache leak would show up on hits (someone else's cached prefix), not on misses where everything is recomputed from your own tokens." — Ryan Tsuji, CTO at airCloset

Then a throwaway account raised the stakes with a receipt.

"One gave us a proper postmortem in which their API gateway was incorrectly handling HTTP 100 status codes… you would receive the response to the prompt that came in before yours" — an engineer who runs LLMs from multiple providers

The Claude Code team closed the loop without closing the question.

"We're confident this is a hallucination but of course take these reports seriously and the team is looking into it." — Thariq, Claude Code team

The unsettling part is not the verdict; it is that from the outside, a hallucination, a local context bleed, and an infra routing bug are hard to tell apart. The reporter guessed a coworker's name but never asked them.

Get breakdowns like this every day

The top Hacker News stories, unpacked — by email, Telegram, or RSS.

Next story · 21 of 22 Read more
🚀Tech

Dan Luu ships AI code with no review, trusting fuzzing over unit tests

Dan Luu merges AI-written pull requests without a human reviewing them, trusting a fuzzing-and-property-testing regime he borrowed from a chip company that shipped under one serious bug a year with no unit tests at all.

★ 164 · 💬 78 · danluu.com · 3 min