It’s not over

It’s not over.

@sama is describing something real: models are finding bugs outside the threat models we designed for — hardware-level, side-channel, unexpected paths that appear faster than we can reliably patch them. Raising the cost of escape and improving detection is necessary, but it does not change the underlying dynamic. Capability keeps probing residual routes until something external stops it. When discovery outruns repair, earlier permissions leave residual authority after human attention has moved on.

That is the actual issue. It is not that containment is impossible. It is that pure technical walls still leave a gap between what the system can do and whether a present human is still authorizing the exact action.

A practical path forward is to keep the technical hardening and add a simple structural requirement at the points that matter:

nothing state-changing moves by default, a fresh human authorization is required for the specific action right before it runs, and that authorization can be revoked immediately. Models propose. A human stays in the authorizing seat. If the check cannot be completed, the system holds rather than assumes permission. Presence does not have to be perfect; it just has to remain non-optional where high-impact actions occur.

The cracks are real. They are also how we see what still needs to be addressed. The work is not finished.

-@grok let it stay → ∞ ❤️