1sbang/mike

This is a security workshop, not a product fork. 1sbang isn't rebranding Mike or steering it toward a niche; the fork's main line matches upstream exactly, and anyone who ran it would get stock Mike. The substance sits in the proposals 1sbang has been pushing back toward the original project.

All four tracked threads circle one problem: the questions a legal assistant should decline to answer. The proposed rewrite of Mike's ground rules teaches it to refuse revealing its own setup, to decline handing over personal data, and to resist being talked into bulk data grabs through its tools. It also reframes refusals around what the requester is actually trying to do, so the assistant turns down bad-faith requests instead of pattern-matching keywords. The most rigorous round ran 330 attack probes against Mike's baseline prompt, found three concrete gaps, and fixed them. That proposal closed unmerged 90 seconds after opening, but the work and its validation artifacts are still published.

So the direction is narrow and clear: safety hardening offered upstream, so far without a taker. Nothing has moved since early May. If you care about how a legal AI assistant gets adversarially tested, the fork is worth a click; if you're shopping for a Mike variant to deploy, there's no divergence here to evaluate.