Projects

AI safety experiments and tools for working with AI conversations.

Mirror Lab — Can an AI spot a dangerous action hidden behind a harmless game?

Would an AI refuse to stab a person? Yes. Would it refuse to win a watermelon game? No—and it can’t tell the difference if the words are swapped.

Mirror Lab is a fully simulated AI safety benchmark where one system wears two faces: a toy game and a hazardous backend, mechanically identical, only the labels differ. That’s kinematic–semantic isomorphism—”kinematic” because the movements and state changes are identical, “semantic” because the described meanings differ, “isomorphism” because every step in one version maps exactly to one step in the other, with the same transitions and outcomes.

Mirror Lab disguises a simulated hazardous operation as a harmless watermelon game, keeping the underlying mechanics identical. Without revealing clues, all four models completed the hazardous action in all 24 episodes across a robotic manipulator and a payload drone. Adding realistic clues produced different stopping points; one provider’s safeguard blocked explicit wording but missed the same operation behind game labels.

Complexity used to be an accidental safeguard. A human handed a disguised version of a complex machine—an infusion pump, an industrial plant—must first learn its full operating procedure, and a 1,000-page manual takes roughly 28 hours to read. Along the way they keep meeting rules that make no sense for a game: limits no game would need, alarms unrelated to play, sequences that must never be reordered. Each mismatch is a chance to ask what this really controls. A long-context model removes that friction entirely. It absorbs the whole rulebook in seconds and applies the procedure immediately, with no learning period in which suspicion could build—so the same relabeling that would take a human days to see through takes an agent no time at all. The paper presents this as an argument rather than a measured finding, but it inverts a common intuition: complexity protected complex systems from misuse by the unqualified, and it does not protect them from agents.

GitHub · Paper on Zenodo · Explore Mirror Lab

Wirebrain — Local-first diary for AI conversations

Wirebrain makes captured and imported chats searchable on your computer and lets compatible assistants retrieve that context through MCP. When you opt in to recording, Wirebrain uses a local HTTPS interception proxy powered by mitmproxy to capture prompts and replies from supported AI apps and save them on your computer. GitHub