[charcircuit]

Stealing a bitcoin wallet by cracking the private key for it also requires red team to be lucky once. Once AI security gets to the point where the probability is infinitesimal for causing actual harm to the business it will be fine.

[_aavaa_]

Yes, and on an infinite time horizon we are all dead.

It’s the time between then and now that we’re talking about.

[charcircuit]

Existing concepts like defense in depth make it exponentially harder for an AI to build a full exploit chain. Even with a full exploit chain with one mistake you'll trigger a detection system which can fool your attack.