Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence β€” Friday

October 30, 2026

πŸ”₯ What Changed

A security-led Friday. A Linux TCP race now has a published remote-root exploit against MPTCP listeners, and Go servers on Linux have listened with MPTCP by default since Go 1.24. The Go team says patch the kernel, with a GODEBUG as a stopgap. A second Go trap: credential RoundTrippers, including golang.org/x/oauth2's Transport, re-add tokens on cross-host redirects. Datalab shipped a faster open OCR model, and Gateway API v1.7 makes HTTPRoute retries Standard. For reading time: PlanetScale on safe planned switchovers.

Also patched: Nomad 2.1.1 (10-29) fixes namespace ACL leaks in the Prefix Search, Dynamic Host Volume read and Scaling Policy read APIs, plus client symlink and log-file hardening. GitHub-reviewed advisories published 10-29 for already-fixed releases: cert-manager GHSA-8rvj-mm4h-c258 (directly created ACME Challenge objects bypass Issuer DNS01 solver policy and use ClusterIssuer DNS credentials; acme-dns credentials can leak; fixed 1.19.6/1.20.3) and five Keycloak advisories fixed in 26.6.4, including GHSA-f5p5-6xmx-p252 (Policy Enforcer authorization bypass by putting the access-denied path in the URL; enforcer fixed 26.0.10). OPA v1.22.0 adds import rego.v2 and tightens bundle signature filename handling (breaking). Go #82472 (GC stretch on high-core amd64): two new CLs, 856483 (lfstack backoff on amd64) and 856521 (P-local buffer pool for stack scanning, βˆ’47% to βˆ’54% on a many-goroutines benchmark on 10 cores; amd64 not measured). Kafka: backport JIRAs for the October CVEs opened against 3.9 (KAFKA-21381 and children; 110985 and 111073 already resolved). MinerU 4.1.0 runs on CPU by default with cross-vendor GPU acceleration and replaces its upload/parse API with /v1/tasks (breaking). AWS API Gateway HTTP APIs in us-west-2 had intermittent invoke errors and latency on 10-29 (19:57–21:17 and 23:17–04:45 IST, worst 03:42–04:15 IST).

Holds: Go 1.27.4/1.26.11 (still not pre-announced; 23 and 21 open issues; #82536 ML-DSA cherry-pick now closed on the 1.27 branch); Go #82717 MPTCP-default proposal (Go Security: premature to change); CVE-2026-98432 Ubuntu/RHEL/AL2 kernels; x/oauth2 CL 854305; Go #82374 x/net/idna (unchanged); #82497 (CL 855561 in review); #82547 unchanged; #82677 (neild cc'd, no CL yet); #82682/#82685 (rsc cc'd); #82472 (CLs above); Kafka 3.9.3 (backports in progress); GitHub RCAs for 10-21/22/26, 10-28 and 10-29 Copilot. Alternates: Google OAuth DPoP-bound refresh tokens (10-15; InfoQ 10-29), SkyPilot Grace Blackwell scheduler, Meta JiTTesting at scale, Cloudflare DDoS Q3 report, Orange Tsai's Edge logic-bug full chain, Mistral Large 4 open weights, Strimzi 1.3.1-rc1, Twisted HTTP/2 rapid-reset GHSA. Closed: #82536 (backport merged), Surya 2.5 (briefed).

🚨 Linux TCP use-after-free CVE-2026-98432 has a demonstrated remote-root exploit via MPTCP listeners, and Go servers on Linux have listened with MPTCP by default since Go 1.24

CVE-2026-98432 (kernel CNA: CVSS 9.8) is a race in IPv4 TCP connection setup: a retransmitted SYN processed on another CPU can build a SYNACK from IP options that the new child socket has already freed (ip_options_build()), and the same bug exists on the request-socket migration path (net.ipv4.tcp_migrate_req=1 or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program). It has been there since Linux 4.4 and is fixed upstream in 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56 and 7.2.10. Anthropic's 10-27 write-up describes a proof-of-concept that gets a root shell, with no credentials, on a Linux server that has a multipath TCP (MPTCP) listener: about 48 minutes of retries to leak a kernel pointer and defeat KASLR, then about 5 minutes to hijack a callback. A lost race only fails the handshake, so an attacker can keep retrying. The Go link: since Go 1.24, net opens Linux listen sockets as MPTCP by default, so practically every Go HTTP/gRPC server on a kernel with MPTCP enabled is exactly that kind of listener. After a user filed #82717 asking to flip the default, the Go Security Team posted a note. It says the best fix is a patched kernel, that the underlying bug is not MPTCP-specific, and that it's "premature" to change Go's default. You can opt out with GODEBUG=multipathtcp=0 at runtime, godebug multipathtcp=0 in the main module's go.mod, or per socket with ListenConfig.SetMultipathTCP(false). Damien Neild adds that dropping IPv4 TCP packets that carry IP options *may* work as a packet-filter mitigation (unconfirmed). Distro status as of this morning: Debian (bookworm 6.1.190-1, trixie 6.12.113-1), Oracle UEK, SUSE Linux Micro 6.2/SLES 16.0 and Amazon Linux 2023 kernel6.12/kernel6.18 have fixes. Ubuntu's linux packages for 22.04/24.04/26.04/26.10 are still "work in progress", and Amazon Linux 2 and the AL2023 default kernel are "Pending Fix".

Why you care: Impact: remote root on internet-facing Linux hosts, with a public, detailed exploit recipe; IPv6 isn't affected by the main path. Who's affected: any Linux box running a Go 1.24+ server (or anything else with an MPTCP listener) on an unpatched kernel β€” that includes most self-built Go services, Kubernetes nodes running Go workloads, and your own VMs. Action: patch and reboot kernels first; on Ubuntu/RHEL/AL2 nodes still waiting, set GODEBUG=multipathtcp=0 on exposed Go services (or sysctl net.mptcp.enabled=0 host-wide, which makes Go fall back to plain TCP), check with ss -Mltn that no MPTCP listeners remain, and consider dropping inbound IPv4 packets with IP options at the edge. Treat those as narrowing the known exploit, not a fix: nobody has shown whether RCE is possible without MPTCP, and the plain-TCP SYN-retransmit path stays racy. No in-the-wild exploitation has been reported.
Read the Go Security Team note β†’

Go HTTP clients that add credentials inside RoundTrip put the token back on cross-host redirects: golang.org/x/oauth2's Transport is still affected, and prometheus/common fixed it without an advisory

net/http strips Authorization when a redirect leaves the original domain, but it does so in Client, *before* the request reaches the RoundTripper. Any RoundTripper that injects a credential on every call therefore re-adds it on each redirect hop and sends it to whatever host the Location header names. vulndb #6887 reports this for github.com/prometheus/common/config (NewClientFromConfig: bearer, basic auth, OAuth2 and configured http_headers, with follow_redirects: true as the default). It was fixed in v0.69.0 (June 2026) with no CVE or GHSA, so govulncheck and Dependabot say nothing; the report notes thanos-io/thanos is still on v0.67.5. The same pattern is open in golang.org/x/oauth2: oauth2.Transport re-adds the bearer token on cross-domain hops (golang/oauth2#805, open since May). The fix, CL 854305 ("send no token on a redirect to another domain"), copies net/http's domain rule and keeps the token off once a chain has left the domain. It has review comments but no maintainer vote yet. The same pattern is also open in cloud.google.com/go/auth (PR #20754, which also covers API keys sent as ?key=) and in google/adk-go's MCP toolset (#1965).

Why you care: This is a common Go pattern (oauth2.NewClient, custom auth RoundTrippers for internal APIs, service-to-service tokens), and it needs no hostile server: a CDN rewrite, a Location built from request input or a lapsed domain is enough, and nothing is logged. Until x/oauth2 ships a fix, set CheckRedirect on clients that carry credentials to refuse (or strip and re-auth on) redirects to a different host, and make your own credential RoundTrippers check req.Response (set on redirect hops) before adding a header. Bump prometheus/common to β‰₯ v0.69.0 in anything that scrapes or remote-writes with secrets. Caveat: the reporter wrote the CLs and PRs too, and no maintainer or CVE has confirmed severity yet.
Read the vulndb report β†’

Datalab Surya 2.5: a 689M open OCR model that reads pages 1.7x faster than Surya 2 on one H100, mostly by writing 2.6x fewer tokens and almost never looping

surya-ocr 0.23.0 and marker 2.1.0 make datalab-to/surya-ocr-2.5 the default. It returns a full page in one pass (typed blocks with positions in reading order, tables as HTML with colspan/rowspan, equations as LaTeX). Datalab's numbers: OmniParseBench (their own open benchmark) 85.9 vs 83.3 for Surya 2, tables 78.3 vs 74.3; olmOCR-bench 83.7 vs 82.8. On one H100 with vLLM 0.25.1 at 192 DPI and 32 concurrent requests: 4.52 pages/s vs 2.72 (Surya 2 with MTP), p95 11.3 s vs 26.0 s, 1,514 vs 3,888 output tokens per page, and 0.1% of requests hitting the token cap vs 3.8%. On an M4 Max with llama.cpp it runs about 0.2 pages/s in 4.3–4.8 GB. Breaking change: surya_table now defaults to HTML mode, and Surya 2.5 has no table row/column boxes (pin SURYA_MODEL_CHECKPOINT=datalab-to/surya-ocr-2 if you need them). Code is Apache-2.0; weights are modified OpenRAIL-M (free for research, personal use and startups under $5M).

Why you care: The speedup comes from shorter output and no runaway loops, which is where self-hosted VLM OCR actually loses money (tail latency and cap-hit retries). That lesson applies to any OCR model you serve. The post is also a good template for benchmarking OCR serving (exact vLLM flags, concurrency sweep, bootstrap intervals). Caveats: every number is Datalab's own; on their multilingual set English slipped from 94.3% to 90.3%, Hindi 69.6% β†’ 67.6% and Bengali 61.6% β†’ 55.3%, which matters for Indian documents, so test on your own pages before switching.
Read the Surya 2.5 announcement β†’

Kubernetes Gateway API v1.7.0: HTTPRoute retries move to the Standard channel, and pre-routing listener filters for external auth arrive as experimental

HTTPRoute retry (codes, attempts, backoff; GEP-1731) is now Standard. attempts defaults to 1 and can be set to 0 to turn retries off per rule. The spec now says implementations should only retry on connection errors "when it is safe to do so". backoff is the base interval, its curve and jitter are implementation-specific, and implementations that retry with no backoff must not claim support for it. New experimental features: pre-routing listener filters (GEP-5224, for external authentication and payload processing before routing, on Gateway and ListenerSet listeners), ClusterTrustBundle references in caCertificateRefs for BackendTLSPolicy and frontend TLS validation (no more copying CA bundles into every namespace), XTelemetryPolicy for tracing, per-address routability, and session persistence moved onto XBackend. Experimental-channel breaking changes cover SessionPersistence field renames and XBackend port fields. The module now requires Go 1.27 and k8s.io v0.37.

Why you care: Retries being Standard means your HTTPRoutes can declare retry policy portably, and that makes retry storms a mesh/gateway concern you can review in YAML. Check attempts and codes against which of your endpoints are actually idempotent before you turn them on. Pre-routing filters are the start of a portable ext-auth hook, so OIDC/JWT checks can sit at the Gateway without vendor CRDs. Implementations (Envoy Gateway, Istio, Cilium, etc.) still have to ship support, so treat this as what's coming over the next couple of releases.
Read the v1.7.0 release notes β†’

πŸ“š Worth Your Time

PlanetScale on planned Postgres switchovers: fix a final WAL position without stopping the old primary, keep synchronous replication valid, and buffer writes at the router

How Neki (PlanetScale's sharded Postgres) moves the primary for pod replacement, kernel upgrades or cell moves. Unlike Patroni, which stops Postgres to get an unambiguous final WAL position, Neki keeps the old primary running so it can roll back fast. Its routers start holding primary-bound queries, and the primary's default for new transactions becomes read-only. Because existing transactions and background work can still produce WAL, Admin samples the primary's position, picks a target that has flushed through it, disconnects the other replicas and records their receive_lsn, takes the maximum, and waits for the target to replay through that point. With synchronous_standby_names = ANY 1 (...), disconnecting the others leaves the target as the only standby that can acknowledge, so every acknowledged commit is on it. Promotion is refused unless at least one replica is attached to the target's synchronous config, so it can't come up blocked on acks. There's a deadline: if the target doesn't catch up, the old primary is made writable again with a config reload. Buffered queries are re-planned against the new topology, so clients see latency, not errors.

Why you care: It pairs well with yesterday's GitHub cutover RCA: the correctness here rests on defining "caught up" precisely (the max of the sampled position and every replica's receive LSN) and on gating the flip on it, with a rollback path that stays cheap because the old primary never stopped. The same moves (fence writes softly, compute a final position, gate on replay, buffer at the proxy, keep a deadline-driven rollback) carry over to any leader handoff you build, whether that's Kafka partition reassignment, a Raft leadership transfer or a dual-write migration.
Read PlanetScale's post β†’