Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence β€” Thursday

October 29, 2026

πŸ”₯ What Changed

A security-heavy Thursday. Kafka published 12 CVEs at once, and one of them lets an unauthenticated client forge a superuser token on OAUTHBEARER brokers that lack a JWKS URL. macOS's bundled git and curl still trust the compromised eMudhra intermediates. Go 1.27 bug reports cover an HTTP/2 flow-control leak, a json/v2 panic and a header-count cap that bit Kubernetes. OpenBMB released an open OCR quality judge. For reading time: GitHub on cutting reads over while replication lagged.

Also patched: go-jose has two more advisories, GHSA-vvm8-gf78-28g4 (JSON parsing panics) and GHSA-4w3g-jfj2-cw2q (CBC-HMAC panic), fixed in v4.1.5. As with 10-28's JWK DoS, v3 (≀3.0.5) has no fix. Echo: GitHub reviewed ten more Echo advisories, most fixed in v5.4.0/v4.16.0 or earlier. Two, GHSA-99jh-6h7p-pp36 (the proxy forwards a client-supplied X-Real-IP) and GHSA-246p-cpwv-v3jq (rate limiter bypass with a spoofed X-Forwarded-For), list no patched v4 version (≀4.16.2), so v5 is the fix. Kubernetes 1.37.2/1.36.6/1.35.10/1.34.13 (10-28) are now built with Go 1.26.10, which carries the eMudhra distrust. cert-manager v1.20.5 removes eMudhra trust (Debian 13 base, Go 1.26.10) and picks up the x/net HTTP/2 fixes. Bun 1.4.4 (10-22) shipped the eMudhra fix, which closes that hold. Docling 2.144.0 adds an opt-in --replace-broken-text-layer: it OCRs only the PDF pages whose embedded text layer decodes as garbage. knative.dev/pkg GHSA-2hw4-c9m4-f2fj: admission webhooks didn't limit request-body size, so a large body could exhaust memory.

Holds: Go 1.27.4/1.26.11 (not pre-announced; 23 and 19 open issues; #82536 cherry-pick approved); Go #82374 x/net/idna (unchanged, NeedsFix); #82497 (fix CL 855561 in review); #82547 unchanged; #82677 untriaged; Go #82472 (the reporter found amd64 wait times inconsistent across hardware, so CL 854865 may wait for a larger change); Kafka 3.9.3 (release under discussion on dev@); GitHub root-cause analyses for 10-21/22/26 and the 10-28 Copilot incident. Alternates: Cloudflare Workers KV Instant (Quicksilver as a product, private beta), Airbnb Chronon near-real-time embeddings, Meta Andromeda, OTel filelog protobuf checkpoints, Apache Qpid Proton C CVEs, MCP Python SDK GHSA-w4fh, OTTL v1. Closed: Bun eMudhra, GitHub 10-27 RCA, MiniCPM-OCR (briefed).

🚨 Apache Kafka discloses 12 CVEs: forged-JWT superuser access on OAUTHBEARER brokers without a JWKS URL, broker file reads by principals with no grants, and pre-auth broker OOMs

Fixes are in 4.2.2, 4.3.2 and 4.4.0 (one is also fixed in 4.3.1). The ones that matter most: CVE-2026-110985 (important; 4.1.0–4.1.2, 4.2.0–4.2.1, 4.3.0): with SASL/OAUTHBEARER and the default DefaultJwtValidator, if sasl.oauthbearer.jwks.endpoint.url is unset the broker silently falls back to ClientJwtValidator, which checks no signature, algorithm or issuer. It accepts "alg":"none" tokens, so an unauthenticated attacker who can reach the listener can claim any principal, superusers included. This is what was left over from CVE-2026-33557. Fixed in 4.2.2/4.3.1/4.4.0, where that configuration now fails at startup. CVE-2026-112375 (important; 3.2.0–4.3.1): an IncrementalAlterConfigs request on a BROKER resource is validated before authorization, and validation resolves ${file:...}/${env:...} through any ConfigProvider, bypassing the CVE-2024-31141 allowlist. Any authenticated principal, even one with no grants, can read files and environment variables the broker can see. CVE-2026-110952 (moderate; 3.1.0 up to 4.4.0): issuer and audience are only checked if configured. This is the 4.4.0 fail-fast change covered on 10-27, now with a CVE. Broker availability: CVE-2026-111115 (tagged-field count in the flexible request header, parsed for ApiVersions v3 *before authentication*; about 20M declared fields exhaust the heap), CVE-2026-111108 (array pre-sizing before authorization; three concurrent ~90 MB ProduceRequests exhausted a 1 GB heap), CVE-2026-114349 (a ~90-byte compressed batch declares a ~2 GiB record; if it lands on a compacted topic the log-cleaner thread dies and compaction stops broker-wide, including __consumer_offsets, across restarts), and CVE-2026-111073 (PushTelemetry decompression). Connect: CVE-2026-116593 (the config validate endpoint returns config-provider secrets in plaintext) and CVE-2026-115644 (admin.listeners skips REST auth extensions). Streams/clients: CVE-2026-111551 (size fields drive allocations, e.g. ListDeserializer; the poison record recurs on restart), CVE-2026-111496 (punctuation overflow hang) and CVE-2026-83525 (FetchSnapshot controller crash, low).

Why you care: Impact: cluster takeover with no credentials (110985), secret disclosure to any authenticated client (112375), and broker crashes from tiny requests. Who's affected: anyone running self-managed Kafka below 4.2.2/4.3.2/4.4.0, especially OAUTHBEARER listeners and Connect REST endpoints. Action: upgrade (4.3.2 also carries the 10-28 tiered-storage data-loss fixes). Today, without an upgrade: set sasl.oauthbearer.jwks.endpoint.url plus expected.issuer/expected.audience on every OAUTHBEARER listener. Note that upgrading alone doesn't close 114349: you also have to set max.decompressed.message.bytes, because the default is still about 2 GiB for compatibility. No exploitation has been reported, but 110985 is trivial to exploit now that it's public.
Read the CVE-2026-110985 notice β†’

macOS's bundled git, curl and python3 still trust the compromised eMudhra intermediates, and macOS 27.2 overwrote hand-applied fixes

Apple has blocked the 12 certificates for the six compromised eMudhra TLS intermediates since 10-19, and emSign ECC Root CA – G3 since 10-23. Each intermediate has two chains: one to emSign ECC Root CA – G3 and one through a Sectigo cross-sign to AAA Certificate Services. That block only covers Safari, Mail and apps that use Apple's networking frameworks. Apple says the command-line tools that ship with macOS and the Xcode Command Line Tools (curl, git, python3, ruby, perl, openssl) verify against /etc/ssl/cert.pem and "don't check Apple's block or whether the certificate has been revoked", and will only be fixed in "a future software update". A third-party report on 10-28 says macOS 27.2 (released 10-26) didn't fix this and replaced hand-edited cert.pem files, which undid the workaround. Apple's interim fix is to remove emSign ECC Root CA – G3 from /etc/ssl/cert.pem, optionally remove AAA Certificate Services too (needed to block the cross-signed path, but it may break some legitimate sites), then re-run the openssl … | grep check after every macOS update.

Why you care: On a Mac, git clone/push over https and any token sent by curl or Python scripts can be intercepted by anyone on the network path who holds a certificate from these keys. The same applies to macOS CI runners. Go programs are a separate case: Go 1.26.10/1.27.3 already reject chains through these intermediates (CVE-2026-111943), so the gap is the system tool bundle, not your Go binaries. Apply the cert.pem edit on your laptop and any macOS runners, and recheck after each OS update.
Read Apple's support note β†’

Go 1.27 field reports: an HTTP/2 client flow-control leak that eventually hangs long-lived connections, a json/v2 Marshal panic, and a 500-header-value cap that broke kube-apiserver impersonation

(a) #82677 (go1.27.3; untriaged, no CL yet): if a request's context is canceled while RoundTrip is still waiting for headers, response DATA that has already arrived is charged against the connection's flow-control window and never returned. In the reporter's repro (8 hedged GETs, keep the first, cancel 7) the whole 1 MiB window leaked in 41 rounds. In production, a long-lived connection used up the default 1 GiB window in one to two weeks. After that, every request on it returned 200 with an empty body and hung, and because PINGs still worked the connection was never closed. (b) #82682: json.Marshal panics (reflect: call of reflect.Value.Type on zero Value) on a typed-nil pointer stored in a custom interface that embeds json.Marshaler. Go 1.26 emitted null. Fix CL 855463 is in review. (c) #82685: invalid UTF-8 now marshals to the raw U+FFFD character rather than the \ufffd escape. It decodes the same, but bytes, golden files and hashes change. For both (b) and (c), GOEXPERIMENT=nojsonv2 restores the 1.26 behaviour. (d) The Kubernetes 1.37.2 changelog (10-28) notes that Go 1.27's default header-value limit (500, via the new Server.MaxHeaderValueCount) made kube-apiserver return HTTP 431 for impersonated or proxied users with about 500 or more groups. k8s.io/apiserver now allows 16,384. Also, #82497 (MaxConcurrentStreams reset to 250) now has fix CL 855561 in review.

Why you care: Hedged and cancel-on-first-response calls over net/http HTTP/2 are a common Go backend pattern, and (a) only shows up after days or weeks, as 200s that never finish. grpc-go has its own transport and isn't covered by this report. Until there's a fix, recycling HTTP/2 client connections on a timer limits the damage. If you own a Go 1.27 HTTP server or auth proxy that forwards large identity headers (groups, impersonation, X-Forwarded-* chains), set MaxHeaderValueCount explicitly. Run your JSON golden tests under 1.27 before you roll it out.
Read Go issue #82677 β†’

OpenBMB MiniCPM-OCR: a 1.3B open document parser plus Jev, a 2B model that scores any parser's output so you know which regions need review

Apache-2.0, weights included. MiniCPM-OCR 1.0 (1.3B) parses Chinese and English documents and PDFs (text, tables as HTML, formulas as LaTeX, charts, seals) and runs on vLLM, SGLang, llama.cpp and Ollama from day one. MiniCPM-OCR Jev (2B) takes an image region and an OCR result and returns a correctness score from 0 to 1. On OpenBMB's own JevOCRBench it reports AUROC 93.6, compared with 90.1 for JEV-35B-A3B and 93.8 for Qwen3.5-397B. Across outputs from four different parsers (MiniCPM-OCR, GLM-OCR, PaddleOCR-VL-1.6, Xiaomi-OCR-0) it stays between 91.9 and 93.4. The new sdk-v1.0.2 adds a minicpmvocr parse CLI for PDF/image β†’ Markdown, with an auto-downloaded PP-DocLayoutV3 layout model, 32 concurrent recognition requests by default, and a Python API over vLLM, OpenAI-compatible endpoints, llama.cpp or Ollama.

Why you care: Jev is the useful idea here: a small, open verifier you can put behind *any* parser, sending low-scoring regions to a retry or a human instead of trusting everything. That's the control most invoice, contract and KYC ingest pipelines lack. With the pip SDK and Ollama backend it's an afternoon experiment. Caveat: every accuracy number is author-run, partly on their own benchmarks, it covers Chinese and English only, and no independent evaluation exists yet.
Open the MiniCPM-OCR repo β†’

πŸ“š Worth Your Time

GitHub's 10-27 RCA: reads were cut over to a new MySQL cluster while gh-ost replication was up to an hour behind, so GitHub App tokens 404'd

GitHub was moving 7 tables between two MySQL clusters (IAD and Azure Central US) with gh-ost. One of them holds installation records for GitHub App server-to-server tokens. A scheduled cleanup job on that same table deleted about 59 million rows (04:44–06:17 UTC), which pushed inter-cluster replication delay to as much as an hour. Reads were then cut over to the new cluster anyway (06:55–07:21 UTC). The new cluster was missing recent installations, so authenticated git and API requests got 404s: 1.24% on average (2.6x normal) and 2.49% at peak (5.3x), until the write cutover finished at 08:01 UTC. GitHub says the impact "began, but was not detected". Fixes: disable overlapping cleanup jobs before migrations, and block read cutover while replication lag is high.

Why you care: It's a short, concrete example of a gap many online-migration runbooks share: they check that the copy *finished*, not that it's *caught up* at the moment you flip reads. Bulk deletes and backfills on the source are what blow up the lag. A lag gate on the cutover step and a freeze on batch jobs during migrations both apply directly to any table or cluster move you run, whether with gh-ost, Debezium-based moves or dual writes.
Read GitHub's incident report β†’