AI has learned to break software. Yours was never in the room.
Anthropic built a model called Mythos that found tens of thousands of exploitable flaws across every major operating system and browser, in code that had survived decades of human auditing, and wrote working exploits for most of them on the first try. Anthropic will tell you themselves it's too dangerous to release. So access goes by invitation, and the invitations went to the biggest technology companies and banks, which are using it right now to harden the most modern software on earth.
But the world doesn't run on the most modern software on earth. It runs on enterprise software nobody gives talks about: internal tools, line-of-business applications, and vendor code that got installed once and now quietly touches everything. Some of it really is ancient, a COBOL job settling paychecks at 2 a.m. on a mainframe older than its operators. Most of it is simply ordinary, a few years past the team that wrote it, running somewhere nobody has needed to look. That software isn't behind the velvet rope; it's in a different building entirely, one no AI lab partner list was ever going to reach.
Legacy system conversions that fail
9 in 10
Replacement is the answer everyone reaches for, and it mostly doesn't arrive. Most of this software will still be running in ten years, close to the state it is in today, and nobody is hardening it.
And the capability is escaping on a schedule. Open-weight models are closing the gap, and every few months they get shrunk and distilled into something that runs on hardware anyone can buy. Frontier-grade exploitation is turning into a download, and a download ends up in the hands of script kiddies carrying capability that used to require a nation-state. What they point it at is whatever is soft and reachable, which is most of what any large organization runs: the internal systems with one maintainer left, and underneath them the banking, power, water, and logistics code that was never designed to survive an adversary that doesn't get tired or bored and reads a million lines of code an hour.
Keikai is being built for everything else.
Part of that is the hard problem, the systems nobody can safely touch. A larger part is simpler: knowing what you are actually running, and deciding what reaches your systems in the first place. The first line of defense is keeping something out. It goes public when it's ready.
Get notified first
You're on the list. You'll hear from us when we surface.
Something went wrong. Please email jason@keikai.ai directly.
No cadence, no newsletter. One email, when it matters.
Or talk to the founder
jason@keikai.ai