Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence — Tuesday

October 27, 2026

🔥 What Changed

A normal Tuesday. Go has a security batch building (an x/net/html fix tonight, 12 private CVEs queued for 1.27.4/1.26.11), plus a fix for 1.27's ML-DSA ClientHello breaking middleboxes. Kafka 4.4.0 ships with an OAUTHBEARER startup check that will break unpinned JWKS setups. Cloudflare explains how a KV-cache pool mismatch leaked other tenants' prompts. Baidu ships a long-document OCR model. For reading time: Wiz on GitHub's Dec 2 pull_request_target enforcement.

Also patched: nats-server v2.15.2/v2.14.9 now respects auth-callout allowed connection types, stops leafnode ACLs from implicitly allowing $MQTT.>, and binds OCSP responses to the cert serial. Valkey 9.1.3/9.0.7/8.1.11/8.0.12/7.2.15 fix a high-severity cluster-mode PUBLISH heap overflow (GHSA-f4q9-j22m-xqq9). Correction to Oct 23: Envoy 1.40.1's distroless image still had the unpatched CA bundle, and 1.40.2/1.39.6/1.38.9/1.37.11 fix it. andybalholm/brotli now has its advisory (GHSA-c9gr-gmg4-2vqv). Reminder: Chrome removes emSign ECC Root CA – G3 today, morning Pacific time (tonight in India). Holds: Kafka 4.3.2 announcement; Go #82374 x/net/idna (may ride tonight's x/net release); grpc-go v1.85; Go backports #82189/#82190/#82297 (headed for 1.27.4/1.26.11); Go #82497 and #82472 unchanged; GitHub root-cause analyses for Oct 21–23 and 26; Google Cloud's final europe-west2 report. AWS says its Sovereign Cloud exercise succeeded but has published no technical write-up, so that hold is closed.

Go patch week: an x/net/html security fix lands tonight, 12 private stdlib CVEs are queued for Go 1.27.4 / 1.26.11, and 1.27.4 drops the ML-DSA ClientHello bit that middleboxes choke on

Three related things. (a) golang-announce pre-announced a golang.org/x/net/html security fix for "US business hours on Tuesday, October 27", which is tonight in India. It covers multiple CVEs. (b) Overnight (02:40–02:54 IST) the Go security team filed 12 private security issues: CVE-2026-97033…97043 as #82567–#82577, and CVE-2026-97054 as #82563. All are release-blockers with backports on the Go1.27.4 and Go1.26.11 milestones. There's no 1.27.4/1.26.11 pre-announcement yet. (c) The same milestone carries #82536 / CL 854185: Go 1.27 will omit ML-DSA from signature_algorithms_cert. That's the fix decided on #81199 (10-26). Since Go 1.27, middleboxes reset or silently drop Go clients' ClientHellos. Reports name Palo Alto (a rule aimed at an old OpenSSL DoS), F5 BIG-IP and Kaseya/Autotask's edge. #82553 is the silent variant: no RST or alert, just a handshake timeout, isolated to extension 50. ML-DSA stays in signature_algorithms, and Go 1.28 drops signature_algorithms_cert entirely.

Why you care: Plan a toolchain bump this week or next. Twelve CVEs in one batch is large, and the x/net fix tonight needs a go get golang.org/x/net@latest in anything that parses HTML. If some partner endpoints give you unexplained TLS handshake timeouts since moving to 1.27.x, this is probably the cause. Until 1.27.4 ships, build those clients with go1.26.10 or cap MaxVersion at TLS 1.2.
Read Go #81199 →

Apache Kafka 4.4.0 is out: OAUTHBEARER brokers now refuse to start without an audience/issuer, the storage format is one-way, and share groups get dead-letter queues

Announced on the mailing lists at 01:22 IST (thread): 25 KIPs and 963 commits. Upgrade traps from the upgrade notes: (1) if sasl.oauthbearer.jwks.endpoint.url is set, the broker's JWT validator fails at startup unless sasl.oauthbearer.expected.audience and expected.issuer are set too. The escape hatches are sasl.oauthbearer.allow.unverified.audience/issuer (default false). (2) KIP-1170 storage directories aren't forward-compatible: the broker must be at least the version of the kafka-storage tool that formatted them. (3) Protocol decoders now reject arrays over 1,000,000 elements and more than 10,000 tagged fields. (4) broker.id is deprecated in favour of node.id, and goes away in 5.0. New features: share-group dead-letter queues (KIP-1191, share.version=2, errors.deadletterqueue.topic.name), transactional state stores for exactly-once Streams (KIP-892, enable.transactional.statestores, off by default), async bootstrap DNS (KIP-909, bootstrap.resolve.timeout.ms), misrouted-connection detection (KIP-1242), unregistering KRaft controllers (KIP-1312, MV 4.4-IV2), and delayed tiered-storage upload (KIP-1241). 4.3.2 is staged on the website but not announced yet.

Why you care: The OAUTHBEARER change is a startup failure, not a warning. Clusters that validate JWTs against a JWKS endpoint without pinning audience and issuer won't come back after a rolling upgrade. Set both before you upgrade, which you should be doing anyway. If you were waiting for DLQs before using share groups (queues) for work dispatch, they're here now.
Read the Kafka 4.4.0 announcement →

Cloudflare postmortem: 155 Workers AI responses leaked other tenants' KV cache, caused by an undersized prefill pool plus an SGLang disaggregated-mode bug

Published at 03:31 IST, coordinated with SGLang. Between Aug 27 and Sep 9, prompts to GLM-5.3 longer than 681,280 tokens (the prefill KV pool) were rejected, because the pool was smaller than the advertised 1,048,576-token context. Each rejected prompt was swapped for a one-token placeholder. In disaggregated prefill/decode mode, both admission queues checked finished() rather than whether the request had been rejected, so the request ran anyway. Decode reserved 8 KV pages, received 1, and attended over 7 stale pages holding other users' prompts. The client got HTTP 200 back. KV pages are not zeroed between requests, and SGLang sizes the pool from free GPU memory at startup, with nothing forcing it to match the context length. Pre-launch tests checked status codes, not output. The bug was reported via HackerOne on Sep 8. On Sep 9 Cloudflare declared an incident at 13:22, set the model private at 14:00 and restored it at 19:40. It found no evidence of deliberate exploitation. Fixes: a ~2.04M-token pool, admission-time rejection in SGLang (#36988, #38935), deferred KV release, per-transfer checksums, and a deployment gate that compares KV pool size to context length.

Why you care: This is a classic distributed-systems bug in an inference stack. Two components disagreed about request state, and shared memory was reused without being cleared. If you or a vendor run disaggregated SGLang, upgrade to v0.5.22 or later. After any change to weights, quantization or memory settings, compare the startup-logged pool size with the context length on every node. The same lesson applies elsewhere: assert on response content, not just HTTP codes.
Read Cloudflare's postmortem →

Baidu Unlimited OCR 2: an MIT-licensed 3B-A0.5B document parser that handles hundreds of pages in one pass with bounded KV cache

Posted to Hugging Face 10-26 (README release note dated 10-27, code). It uses Sparse Reference Sliding Window Attention plus "unlimited-horizon" decoding. All numbers are Baidu's own, measured on an A800-80GB. Peak KV cache for a hundred-page document drops from 47.00 GB to 1.77 GB (batch 1), and throughput rises from 360 to 1,315 tokens/s (batch 16). Per step, that's 7.0× faster than full attention and 5.6× faster than v1's R-SWA. Scores: MPDocBench 84.59; 77.06 on MarathonDocBench, Baidu's own new 100–300-page benchmark, where Kimi-K3 scores 33.74; and OmniDocBench 94.02.

Why you care: This is the first open parser pitched at whole long PDFs rather than one page at a time, and its memory profile makes that practical on one GPU. Caveats: every benchmark is self-run, partly on Baidu's own benchmark; heading structure is weak on long documents (HeadingTEDS 48.20); it needs a custom SGLang build; and it's less than a day old. Worth a bake-off on your longest documents, but not yet a replacement for page-level pipelines.
See the Unlimited OCR 2 model card →

📚 Worth Your Time

Wiz, "The Pwn Request Reckoning": what GitHub's 2026 pull_request_target lockdown fixes, what it leaves open, and the Dec 2 deadline

Published 10-26. A timeline of the changes: actions/checkout v7.0.0 blocks fork checkout in untrusted triggers (06-18) and is backported to v2.8.0/v3.7.0/v4.4.0/v5.1.0/v6.1.0 (07-20); read-only cache tokens (06-26); cache-mode GA (09-10); Workflow Execution Protections GA (09-17). On Dec 2, GitHub starts enforcing the default event policy that blocks pull_request_target in public repos without an explicit policy (GitHub changelog). Private and internal repos and workflow_run aren't covered. Still exploitable: explicit opt-ins, SHA pins to pre-fix checkout versions, issue_comment bot commands, and script injection. It ends with an audit checklist.

Why you care: If any of your public repos use labelers, preview deploys or fork CI with secrets, they will either break on Dec 2 or need an explicit allow policy, and every explicit allow is a workflow attackers will look for. Check Policy Insights now and bump any SHA-pinned actions/checkout refs to the patched minimums.
Read the Wiz post →