grpc-go (merges Sep 24) — The P0 transport deadlock (regression since v1.82.1: concurrent unary RPCs on one ClientConn when socket send buffers are constrained) now has merged fixes on master, milestone 1.85 Release. #9431: stop client-side read throttling based on controlBuf size — server can also throttle, so mutual throttle → permanent deadlock. #9433: refine server throttling to only count peer-induced control/reset frames, and raise default GRPC_GO_EXPERIMENTAL_CONTROL_BUFFER_THROTTLE_LIMIT from 100 → 10,000. Issue #9330 remains open (PRs say “Addresses”, not “Fixes”); no release yet (latest stable still v1.84.0). Material update vs the 2026-09-16 issue brief.

Meta Engineering (Sep 23) — Extends Meta’s confidential-computing runtime (Private Processing / TEEs) from WhatsApp & Meta AI chats into AI glasses so hyper-personalized cloud inference stays cryptographically unread by Meta, host OS, and hypervisor. Five hard requirements: hardware isolation (SEV-SNP–class CPU+GPU CVMs), fail-closed + public verifiability (append-only transparency ledger), non-targetability (blind-signed credentials + OHTTP relays), RA-TLS attestation, and encrypted storage with user-provided keys. Stateful-memory hard part: encrypting at the device and querying a normal cloud DB leaks access patterns — so Meta moved the storage engine inside the TEE. Ops: aggregate-only observability.