Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence β€” Tuesday

November 10, 2026

πŸ”₯ What Changed

A normal Tuesday with four items. Go 1.27.4 / 1.26.11 is not out yet: it was pre-announced for US business hours today, so expect it tonight IST, and go.dev/dl still lists 1.27.3 / 1.26.10 as of 08:40 IST. Separately, the Go team opened a new PUBLIC-track CVE (CVE-2026-97047, #82948): the HTTP/2 server dedups trailers in O(NΒ²), so a ReverseProxy in front of a malicious backend can be pushed into a CPU DoS. It is not one of the 12 pre-announced CVEs. OpenSSH 10.6p2 ("sshole") fixes a password-auth bypass on musl systems without PAM, which includes Alpine's default openssh-server; musl said on 11-09 that crypt() may return NULL on failure in its next release cycle. AWS launched Nitro Sandboxes for EKS, which runs each pod in its own Nitro-isolated VM via a RuntimeClass. Istio 1.32.0 is out with AuthorizationPolicy on HTTPRoute, ORCA utilization-based load balancing, and required JWT rules.

Also seen: MIT krb5-1.22.3 (11-09) fixes 21 CVEs. The serious ones are a PAC server-checksum logic error (CVE-2026-111423) plus unkeyed CAMMAC checksums (CVE-2026-71199), which combined with a KDC authdata-filtering loophole could let fake PAC/CAMMAC values reach application servers. Also kadmind lockdown_keys (CVE-2026-107884), RPCSEC_GSS service-type confusion (CVE-2026-118380), and AUTH_GSSAPI now disabled by default. Go #82930: recover in a range-over-func body returns nil after a stack copy (go1.26.10/1.27.3/tip; CL 861541). Zitadel v4.20.2; OpenTelemetry Collector contrib v0.165.1. NATS 2.15.3/2.14.10 still RC.2.

Holds: Go 1.27.4/1.26.11 (pre-announced for US business hours Tue 11-10; still not on go.dev/dl or golang-announce at 08:40 IST, so expect it tonight IST; check whether CVE-2026-97047 rides along); upstream Keycloak fix for CVE-2026-120984 (latest upstream still 26.8.2); gopls release for #82821 + #82891 (both fixed at tip; latest release still v0.24.0); CL 858982 (#82853) still open; Go #82945 TLSNextProto regression awaiting triage; CVE-2026-98432 OpenShift CoreOS; x/oauth2 CL 854305; Kafka 3.9.3 (code freeze 11-20); NATS 2.15.3/2.14.10 GA; Postgres 19 RC2 ~11-12 / GA ~11-19.

Go CVE-2026-97047: HTTP/2 server trailer dedup is O(NΒ²), so a ReverseProxy can be driven into CPU DoS by a malicious backend

Damien Neil (neild) filed #82948 and labeled it Security + release-blocker, milestone Go1.28. The HTTP/2 server dedups the contents of the Trailer header it sends with a linear scan for each trailer, which is O(NΒ²) in the number of unique trailers. That is harmless for normal handlers, but a ReverseProxy relaying a response from a malicious backend can be made to forward a very large number of trailers and burn CPU, an effective denial of service. It is designated PUBLIC track with CVE-2026-97047. Fix CL 862201 ("net/http/internal/http2: avoid O(N^2) in trailer deduplication") is in review. It is not among the 12 CVEs pre-announced for today's 1.27.4/1.26.11 (97033–97043, 97054).

Why you care: Go Γ— Microservices. It hits API gateways, sidecars, and BFFs built on httputil.ReverseProxy that front backends you don't fully trust (multi-tenant or user-supplied upstreams). When tonight's minors land, check whether the release notes include CVE-2026-97047 or whether it waits for the next point release. Related: #82945 reports that a TLSNextProto["h2"] handler installed after x/net/http2.ConfigureServer is silently ignored since go1.27, which breaks gRPC+HTTP/2 single-port muxers such as elastic/gmux; it has no maintainer response yet.
Read Go issue #82948 β†’ Also: go.dev β†’ Also: github.com β†’

OpenSSH 10.6p2 "sshole": password-auth bypass on musl without PAM (Alpine default sshd); musl may switch crypt() to NULL-on-failure

OpenSSH 10.6p2 (11-05) fixes a critical password-authentication bypass in sshd on Linux systems built and run with PAM disabled and using musl libc. Alpine is named as affected, and the reporter says Alpine's normal openssh-server package is built without PAM. For any account whose password field is exactly * (key-only users, root, service accounts), a password of 257+ characters makes musl crypt(3) return *. sshd then compares that to the stored * and lets you in. Per the reporter, this requires PasswordAuthentication yes (the default) and at least one other account with a real $5$/$6$ hash, which sshd borrows as a salt. The upstream fix (commit 4353fdc4) refuses to compare stored fields shorter than MIN_PASSWORD_HASH_LEN. The mitigation is PasswordAuthentication no. New on 11-09: musl maintainer Rich Felker said crypt() "will possibly be changed (to always return null for failure) in the next release cycle", but not as a build option and not backported to stable distros, because it may break callers.

Why you care: Auth Γ— Cloud. Alpine-based bastions, jump hosts, appliance images, and "debug" containers that run sshd are exactly the affected configuration, and nobody normally thinks of key-only accounts as reachable by password. Upgrade OpenSSH (or the Alpine backport) and restart sshd, or disable password auth. We missed this on 11-05, so it is reported now alongside the musl follow-up.
Read oss-security post β†’ Also: www.openssh.org β†’ Also: www.openwall.com β†’

AWS Nitro Sandboxes for EKS: each pod in its own Nitro-isolated VM via runtimeClassName: isolated-sandbox (plus a confidential mode with KMS attestation)

AWS launched Nitro Sandboxes for Amazon EKS. A new EC2 capability lets the Nitro Hypervisor carve a hardware-isolated slice of CPU and memory (a "Nitro Partition") out of a running instance, and the container runtime boots a lightweight VM there for each sandboxed pod. Kubelet still sees a normal pod: exec, logs, port-forward, probes, EKS Pod Identity, CSI volumes, and VPC CNI all work, and AWS says it passes every Kubernetes conformance test. There are two RuntimeClasses. isolated-sandbox protects the host and neighbouring pods. confidential-sandbox treats the operator as untrusted: the Nitro Security Module attests the booted image and Cedar policy, and AWS KMS releases secrets only to a matching sandbox, which AWS says stops a cluster admin from reading pod memory. Shared node resources (network, disk I/O, DNS, logs, conntrack) are throttled by default, and a sandbox.amazonaws.com/network-isolated annotation boots the VM with loopback only. It requires Kubernetes 1.36+. Auto Mode ships both RuntimeClasses, and standard EKS needs a sandbox-enabled AMI. AWS says it integrates with the SIG agent-sandbox project, whose warm pools give about 2 s claims, and with Temporal.

Why you care: Cloud/K8s Γ— Auth/IAM. This is a managed VM-per-pod boundary for untrusted code (agents, CI builds, multi-tenant SaaS) without a dedicated node per tenant. It follows the 11-05 OpenAI LPE report, which recommended gVisor/Kata/Firecracker-style isolation. The isolation, conformance, and admin-memory claims are vendor statements; AWS points to a deep-dive post to follow.
Read AWS announcement β†’

Istio 1.32.0: AuthorizationPolicy can target HTTPRoute, ORCA utilization-based load balancing, required JWT rules, one east-west gateway for mixed ambient/sidecar meshes

Istio 1.32.0 supports Kubernetes 1.33–1.37. Highlights: AuthorizationPolicy targeting Gateway API HTTPRoute via targetRefs (experimental), so ALLOW/DENY rules apply only to traffic matched by that route, with a new IST0178 analyzer warning when an EnvoyFilter inserts an HTTP filter alongside one; RequestAuthentication jwtRules can be marked REQUIRED, rejecting requests missing that token even if other rules pass; PILOT_DISABLE_SELF_SIGNED_CA makes istiod refuse to start a self-signed CA; a new backendUtilization DestinationRule policy uses backend-reported ORCA load reports to send less traffic to busier endpoints; outlier detection gains outlierDetectionHttpErrorCodes and failurePercentageThreshold; defaultInboundHttpRetryPolicy configures or disables ({}) the default sidecar inbound retries; a single istio-east-west gateway can serve sidecar multi-network traffic on 15443 (needs AMBIENT_ENABLE_MULTI_NETWORK); GOMAXPROCS is removed from all charts because Go now derives it from CPU limits.

Why you care: Microservices Γ— Auth/IAM. Route-scoped authz and required-JWT close two common gaps in mesh policy. ORCA-driven load balancing and failure-percentage ejection matter for gRPC fleets. 1.30/1.31 users still need the 11-06 security patches (ISTIO-SECURITY-2026-009) whatever they decide about 1.32.
Read Istio 1.32 announcement β†’ Also: github.com β†’

πŸ“š Worth Your Time

Nothing exceptional enough today.