What the gate checks: lend's approvals, Starlette's Host header, FastAPI's Depends object
Four of today's items describe controls, and each one is precise about what the control does not cover.
lend asks for a Touch ID approval before a command from a remote coding agent runs on the laptop, and its policy denies everything not listed. Its author still writes that approval decides what runs, not what the command then does: an approved shell or interpreter reads further instructions from stdin after the approval, and the caller details in the dialog come from the VM and can be faked.
The scans SANS ISC recorded against LiteLLM work because an earlier control trusted an input it never checked. Starlette before 1.0.1 built the URL from the Host header and used that URL for access decisions; the attackers now put brackets into the header, apparently to get past filters written around the public exploit. Johannes Ullrich's advice is to check the input - allowed hostnames - rather than the payload.
FastAPI's new max_concurrency limit is attached to a Depends() object. Reuse the object and routes share a limit; write a new one per route and each gets its own, though the function is the same. The limit is per worker process and per event loop, and the documentation shows how a request can wait on itself.
Kubernetes' account of the IPVS backend fits the same pattern from the other side: it solved the two scaling problems it was built for, but the kernel API it relied on could not express every case the Services API needed, and the missing cases were never closed.

Why it matters
In each case the boundary is in the source, written by the people who built the control. Reading that sentence before relying on the control is cheaper than discovering it from an incident.