The report
A workflow log is evidence of an executed machine-learning task, but it may also contain resource paths, input details and operational secrets. One Kubeflow Pipelines advisory describes a log endpoint that could return a run's output without confirming that the caller was authorized for the requested run. An attacker would still need to reach the API and identify a target run. The public notice does not demonstrate secret theft from a named deployment.
A second advisory describes the cost of processing untrusted upload archives. Decompression could exhaust a shared server's memory before all later checks finished. This is an availability failure rather than a bypass of another tenant's log permissions. Both involve a shared ML execution service, but their triggers and impacts differ.
The cross-run log Case and the archive expansion Case keep those conditions separate. Authorization should bind every log request to a principal and a run, while upload parsing should enforce decompressed-size and resource budgets before expensive work begins.