Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence — Friday

November 13, 2026

🔥 What Changed

A normal Friday with four items. HashiCorp Vault 2.2.2 fixes an ACL hole: explicit deny rules did not apply to case variants of the denied path, so a deny on secret/data/example could be sidestepped with secret/data/EXAMPLE. runc 1.5.3 / 1.4.4 / 1.3.7 fix CVE-2026-119101 (high), where a malicious image can make runc restore bind-mount an arbitrary host directory read-write. This only matters if you checkpoint and restore untrusted containers. A new Go compiler wrong-code bug (#83036) reproduces on every release from go1.23 to tip: a loop-carried bool is silently rewritten into the loop's entry condition. For Worth Your Time, LinkedIn explains why it replaced Argo CD with a library-based, per-deployment GitOps engine across 250+ Kubernetes clusters.

Also seen: Echo v5.4.5 / v4.16.4 fix an open redirect in the HTTPS/WWW redirect middleware (GHSA-3f6j-cw5h-qw9c); RealIP() now strips IPv6 zones. Go #82374 (x/net/idna Punycode CPU DoS, CVE-2026-97044) got 1.27/1.26 backport issues #83034/#83033. Go #83037: a tip-only regression (CL 859161) where a panic in a TLS GetConfigForClient or ConnState hook crashes the whole server. Go #83024 HTTP/2 client conn hangs forever if its read loop panics. OPA v1.22.4 rebuilt on Go 1.27.4. Kyverno 1.20.0 GA. Docker 29.10.0-rc.2. Redis 8.12-rc3. Kafka KAFKA-21468 (SyncGroup COORDINATOR_LOAD_IN_PROGRESS unhandled by classic clients) merged.

Holds: Go 1.27.5/1.26.12 (watch for CVE-2026-97047 #82948 and CVE-2026-97044 #82374 backports); Go #83036 fix CL; Go #82945 TLSNextProto h2 regression; Go #82977/#82978 windows backports; Keycloak CVE-2026-120984 upstream fix; gopls release for #82821 + #82891; CVE-2026-98432 OpenShift CoreOS; x/oauth2 CL 854305; Kafka 3.9.3 (code freeze 11-20); Postgres 19 GA ~11-19; Cloudflare Meerkat follow-ups.

HashiCorp Vault 2.2.2: explicit ACL deny rules could be bypassed with case variants of the denied path

The changelog's SECURITY entry is core/acl: harden explicit deny rule evaluation by matching requested paths case-insensitively. Before this fix, a policy that denied secret/data/example did not deny a request for secret/data/EXAMPLE. Explicit deny rules now match case-insensitively, subject to the usual rule-specificity order. The release also moves Vault to Go 1.27.4, which picks up Tuesday's toolchain CVE fixes. It updates the auth plugins (jwt v0.27.2, kerberos, cf, oci, alicloud) and the database/secrets plugins. Agent and Proxy also stop rendering interactive HTML returned by custom plugins. No CVE or HCSEC bulletin was linked in the release notes at scan time.

Why you care: Auth/IAM is a rising board category, and "deny wins" is the guardrail many Vault policies depend on, for example allow secret/* but deny one sensitive path. If your secrets engine treats paths as case-insensitive, the deny may never have held. Upgrade and audit policies that rely on explicit deny.
Read Vault v2.2.2 release →

runc 1.5.3 / 1.4.4 / 1.3.7: CVE-2026-119101 lets a malicious image make runc restore bind-mount a host directory read-write

CVE-2026-119101 (high): crafted symlinks in an image's rootfs confuse how runc names external mounts for CRIU. As a result, runc restore can bind-mount an arbitrary host directory, read-write, into the restored container. You are affected only if you both checkpoint and later restore containers running untrusted images. Checkpoint/restore needs root, so rootless runc is not affected. The same bug class was fixed in crun (CVE-2026-115520) and in CRIU itself (CVE-2026-119614). The release also stops runc restore from temporarily making read-only bind mounts writable. Other fixes: containers without their own mount namespace no longer change the propagation of the host's /; root.readonly no longer clears nosuid/nodev/noexec/nosymfollow on the rootfs; and the defaults option no longer breaks tmpfs mounts. 1.6.0-rc.2 was cut at the same time.

Why you care: Kubernetes is your #3 weight. Container checkpointing (the kubelet checkpoint API used for forensics and live migration) is the path where this matters, and the CVE's impact is host filesystem write. Even without C/R, take the mount-propagation and readonly-flag fixes when your container runtime picks up this runc.
Read runc v1.5.3 release → Also: github.com →

Go #83036: compiler phiopt rewrites a loop-carried bool into the loop's entry condition, producing wrong results from go1.23 through tip

In a short reproducer, a for w := !first; ; w = true loop sets refusing = true after the first iteration. The program should print [false true true] but prints [false false false] with default flags. The refusing value carried across iterations is replaced by the entry-branch condition through a phiopt special case. The reporter reproduced it on go1.23.12, go1.26.4–1.26.11, and go1.27 and tip on the playground, on both arm64 and amd64. The issue is untriaged, with no CL yet. It follows the run of fuzzer-found wrong-code bugs from earlier this month (#82825 and others).

Why you care: Go is your top weight. This is silent miscompilation in an ordinary pattern (a state flag in a retry or wait loop), so there is no crash and no race detector hit, just wrong control flow. There's nothing to do yet except watch for a fix and 1.27.x/1.26.x backports. If you ever see a loop flag that "never flips", this is the bug to suspect.
Read Go issue #83036 →

📚 Worth Your Time

How LinkedIn scaled deployments with a custom GitOps engine, replacing Argo CD across 250+ Kubernetes clusters

LinkedIn runs 250+ clusters (up to 5,000 nodes each), 4,000+ services and over 100,000 deployments a week. It started with one Argo CD instance per cluster behind an orchestrator and hit seven problems. Argo CD has no observedGeneration, so sync and health status couldn't be tied to a specific deployment. Stale repo-server cache could resolve the wrong revision. One watch per GVR shared by all apps made sync time grow with pod count (under 3 min at 60k pods, over 10 min at 80k). Server-side-apply managedFields get truncated after 10 field managers, so a container removed from a manifest could survive on the live Pod. CRD changes needed fleet-wide controller restarts. The new engine is a library, not a controller: one ephemeral pod per deployment with no shared state, PUT not PATCH (retried on resourceVersion conflicts with backoff), a 3-way diff for first-class drift detection, synchronous convergence tracking and API-server error passthrough. According to InfoQ's summary, deployment-infrastructure failures fell from 925 to 63 despite 3.7× volume, and median time to detect a failed deployment dropped from 16.1 to 2.45 minutes. Argo CD is still used for cluster add-ons.

Why you care: This covers Kubernetes (#3) plus distributed-systems design: noisy-neighbour shared watches, SSA ownership ambiguity and controller versus per-request isolation. Two checks you can run today without a rewrite: count field managers on long-lived objects, and see whether Argo sync latency tracks total pod count.
Read LinkedIn GitOps engine post → Also: www.infoq.com →