Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence — Wednesday

November 4, 2026

🔥 What Changed

A quiet Wednesday with three items, all in What Changed. The same AI-assisted fuzzing campaign that filed ten Go compiler bugs on Monday has now found one on amd64: since at least Go 1.24, a shift combined with a comparison against the shift width can return the wrong boolean. Docker Engine 29.9.3 fixes a Swarm worker-to-manager certificate escalation and lost overlay encryption after a firewalld reload. A gopls report shows the CVE-2026-42503 fix missed the -debug flag, so a cloned repo's VS Code settings can expose your open source files to the LAN; the fix is merged but not released. Nothing in Worth Your Time cleared the bar today.

Also seen: Caddy v2.11.8 (11-04 06:46 IST) is a large bug-fix and security-hardening release with breaking changes. It rejects invalid or dots-only hosts with 400 on every protocol (HTTP/2 and HTTP/3 used to accept hosts containing / or \), makes request_body max_size reject on Content-Length up front, fixes client_auth and per-listener TLS policy overwrites, and makes reverse-proxy timeouts behave. Argo CD v3.6.0 (11-04 03:59 IST) is GA, with ApplicationSet progressive-sync metrics and conditions. k3s v1.37.2+k3s1/v1.36.6+k3s1 shipped (11-03). InfoQ summarized LinkedIn's OpenTelemetry client-to-backend tracing: a pageForestId grouping above traceIds, coordinated head sampling keyed on a pseudonymised member id, and Kafka into ClickHouse (secondary source; no LinkedIn primary post found). Prometheus 3.16.0-rc.4.

Holds: Go 1.27.4/1.26.11 (go.dev/dl still 1.27.3/1.26.10; now also carrying CL 858262 for #82825, CL 857344 for #82777 and #82771); Go #82717 MPTCP default (no comments since 10-31 21:15 UTC); CVE-2026-98432 Ubuntu -updates, Ubuntu cloud kernels and OpenShift; gopls release with CL 858103; x/oauth2 CL 854305; go-jose v3 fixes; Kafka 3.9.3; Go #82785 review; moby/swarmkit GHSA-628g and GHSA-6jhp publication; Dapr CVE/GHSA IDs; Postgres 19 RC2 11-12 / GA 11-19. Alternates: Netflix AI agents for performance engineering (ByteByteGo, secondary), Grafana trace-to-database-query linkage, Grafana Alloy OTel engine public preview, MinerU 4.1.5/4.1.6, Docling 2.146.0, LWN on LLM reviews for Debian's NEW queue.

Go amd64/386 miscompile: a shift and a comparison of the shift count against the width can share CPU flags that an SBB instruction overwrites, so y == 64 compiles as y >= 64 (go1.24.13 through 1.27.3)

s-celles (same AI-assisted metamorphic fuzzer as Monday's batch) reports that the x86 ops SBBQcarrymask/SBBLcarrymask, which the compiler uses to saturate unbounded shift counts, expand to an SBB instruction that writes the flags register but are missing clobberFlags: true. When the program also tests the shift count against the width, flagalloc reuses the CMPQ $64 flags for the SETcc, and the scheduler can place the SBB in between. The reproducer r := x << y; if y == 64 { eq = 1 }; return r + eq returns 1 for f(1, 65) instead of 0. Other affected shapes on amd64: y != 64 (compiled as y < 64), signed and unsigned x >> y with y == 64, and the 32-bit x << y with y == 32; 386 also miscompiles a decomposed 64-bit comparison. GOAMD64 v1/v2/v3 all behave the same, and source order matters (it's correct when the SETcc is scheduled first). The report says compiler-inserted bounds checks are not affected and no out-of-bounds access was seen; the impact is wrong boolean results. Fix CL 858262 ("mark the x86 SBB carry-mask ops as clobbering flags") is in review, and Jorropo opened 1.26 and 1.27 backports. The same reporter filed #82834 (arm bits.Sub64 wrong borrow, root cause shared with #82786) and #82839 (prove takes an impossible branch through a sign extension).

Why you care: Monday's batch was mostly arm64 and niche architectures; this one is on amd64, where most Go services run. The trigger is narrow (a variable shift count also compared against exactly 32 or 64 in the same function), but it's the kind of code that shows up in bit-packing, varint, hashing and bitmap helpers, and the result is silently wrong rather than a crash. There's no compiler flag to avoid it, so the practical action is the same as Monday: take Go 1.27.4/1.26.11 promptly when they ship, and if a bit-manipulation test fails only on amd64 with a width comparison nearby, check #82825 first. Expect more of these; the fuzzer is now running a --late lowering campaign.
Read Go issue #82825 →

Docker Engine 29.9.3 fixes a Swarm worker-to-manager certificate escalation, overlay networks losing encryption after firewalld flushes Docker's iptables rules, and two BuildKit bugs

The release lists four CVEs. CVE-2026-115539 (SwarmKit, GHSA-628g-hv7f-pgjc): an attacker holding a valid worker-node TLS certificate and key could obtain a manager-node certificate shortly after a Swarm leader change. CVE-2026-115538 (GHSA-6jhp-7446-hf7f): Swarm overlay networks lose their encryption and anti-spoofing protections when firewalld flushes Docker's iptables rules. CVE-2026-116587: excessive CPU when parsing crafted RSA keys (this is go-jose GHSA-mgcc-cww5-fp2q, covered on 10-28). CVE-2026-116591: a possible panic when decompressing zstd data (GHSA-fpmh-frq8-7rj8). The bundled BuildKit v0.34.2 fixes GHSA-w26x-xg27-7qjc (medium; a client with access to the BuildKit control API can set xattrs, including trusted. and security. ones, on host files outside the state directory) and GHSA-2fqp-xg6f-v7gj (Git source resolver cache poisoning can crash the daemon). The two moby/swarmkit advisories weren't published on GitHub yet when checked, so details are only what the release notes say. It also fixes the daemon hanging on container, image and system API requests under load (moby/moby#54041).

Why you care: If you run Swarm, a compromised worker becoming a manager is full cluster control, and the firewalld issue means "encrypted overlay" may have been quietly unencrypted on RHEL-family hosts. For Kubernetes-only shops, the relevant part is BuildKit: any shared or multi-tenant build service (CI runners, in-cluster BuildKit) should move to v0.34.1+. Upgrade build hosts and Docker nodes to 29.9.3.
Read Docker 29.9.3 release notes → Also: github.com →

The CVE-2026-42503 gopls fix missed -debug: gopls -debug=:PORT still binds all interfaces and serves your open files, pprof and RPC traces without auth, and a repo's .vscode/settings.json can turn it on

CVE-2026-42503 (#79211) was fixed by making gopls -listen reject an address with no explicit host. Reporter Sanu1999 shows the -debug flag goes straight to net.Listen("tcp", addr), so -debug=:47781 listens on 0.0.0.0. The unauthenticated debug server exposes open buffer contents (/session/N, /file/{session}/{id}), cache and metadata, RPC traces, metrics and /debug/pprof/* (profile and trace endpoints can also burn CPU). The vscode-go angle: go.languageServerFlags has no restricted scope, so a repository can ship {"go.languageServerFlags": ["-debug=:47781"]} in .vscode/settings.json, and once you trust the workspace gopls starts normally while exposing the debug server to anyone on your network. CL 858103 ("reject implicit INADDR_ANY in debug address") merged on 11-03, but the latest gopls on the module proxy is still v0.24.0 (10-14), so no released gopls has it.

Why you care: This is a developer-laptop leak rather than a server bug, but it's cheap to check. Until a gopls release ships, don't trust a workspace from an unfamiliar repo without reading its .vscode/settings.json, and if you use -debug yourself, give it localhost:PORT.
Read Go issue #82821 →

📚 Worth Your Time

Nothing exceptional enough to recommend today.