scan the image you actually deploy
i want the digest of the image that's going to run and the scan result attached to it. a clean source scan only answers part of the question.
writing
model rollouts need a rollback path
four replicas, four gpus, and a rollout that wants a fifth slot. the release plan needs to account for the hardware as well as the model.
capacity planning for llm serving
a long document with a short answer and a short prompt with a long answer both count as one request. that hides most of the capacity problem.
rag needs an authorization boundary
remove someone's project access, then let them ask yesterday's question again. the answer tells you how well permissions reach the index and caches.
prompt injection is a permission problem
following a malicious instruction from a support ticket through to a proposed email, and the permission checks that should stop it.
the tax files were in the tickets
the ey reporting raises a question about the support workflow: why were full client tax files attached to tickets, and who could retrieve those copies?
read article ↗