Skip to content
senn-techsenn-tech
← IT News
Week 2026-W472026-11-16

IT News Wk 47: Harness Swaps, OpenVDDK, EPYC Venice

Coding agentsHarnessOpenVDDKEPYC 9006PostgreSQL 14

This week the coding-agent story sits in the shell around the model, not in the model. An open harness leaderboard kept the weights fixed and swapped only the program. The open weighting moved from 35.5 percent of tasks solved to 52.0 percent, and cost per solved task fell from $3.35 to $0.45. Virtualisation is the second theme: OpenVDDK replaces the library Broadcom pulled from its download pages at the end of August. It carries migrations today, with one open defect on ESXi 7.0 GA. Then AMD reports EPYC 9006 "Venice" with 256 cores, 16 memory channels and an open question about what the advertised bandwidth costs in procurement. Label key: verified (two independent sources read), reported (one source read, even where we calculated figures from it), vendor figure, derived (our reading). Each entry says which source we actually opened.

On the week's output: we published 40 posts, 25 of them under “From the blog” (including the 3 deep dives, below in full) and 15 under “More from the week”. Many of them contain our own measurements on our systems (gateway, fleet, mail host, price lists); where they do, the post names the place, the date and the method. The digest below is news, not measurement, except where an entry says “measured by us”. The version numbers in the digest were taken on 15 November 2026; each entry states how, and on how many hosts, we queried them.

Deep Dive: Coding harnesses: the same model, up to 26 points apart

On 9 November 2026 the open Harness Leaderboard on Hugging Face replaced OpenCode with version 2.2.6 and recounted the results of the previous version 1.19.0. In the open weighting across all models, OpenCode had stood at 35.5 percent and $3.35 per solved task, rank 8 of 9 as the board displayed it (today it lists ten harnesses, and the old values can no longer be retrieved). After the swap it stood at 52.0 percent and $0.45, rank 1, together with four other harnesses inside the error bars. The model was identical in both runs. A harness is the program around the model. It builds the prompt, proposes tools, runs the calls and decides what stays in context for the next pass.

What the board delivers

The tasks come from Terminal-Bench 2.1 (89 upstream, of which the board scores 87) and DeepSWE (113 upstream, 111 scored). The table shows the state of 15 November 2026: we loaded the machine feed api/index on 15 November 2026 at 16:47 UTC and calculated the share of solved attempts per harness and model ourselves (passed divided by scored attempts, across both benchmarks and all effort levels). The payload (version 0.1.7) counts 39,122 trials and roughly $16,800 in model cost; the harness-comparison post quotes the load of 15:51 UTC, and the values agree. Reported.

Harness (version on the board)GLM-5.3-FlashDeepSeek-V4.1-FlashFlintRipple-65B-PreviewCost per solved task
OpenCode 2.2.661.8 %70.2 %21.2 %52.9 %$1.11
mini-swe-agent 2.4.661.8 %67.8 %20.2 %45.5 %$1.41
Kimi Code 2.2.059.8 %68.4 %19.1 %55.7 %$1.37
DeepSeek Harness 0.2.2-rc.160.2 %70.2 %19.2 %45.6 %$1.31
Hermes Agent v0.22.159.6 %61.4 %21.5 %51.3 %$1.36
Claude Code 2.1.30061.4 %68.9 %15.8 %46.5 %$1.32
Pi 1.2.061.4 %65.1 %16.7 %44.6 %$0.76
Qwen Code 0.25.254.7 %65.0 %11.3 %41.1 %$1.15
Codex 0.162.145.4 %68.1 %16.2 %55.7 %$1.31
OpenClaw 2026.11.135.8 %53.5 %5.6 %20.4 %$4.46

How to read it: the share of solved attempts differs between the best and the weakest harness by 26.0 points on GLM-5.3-Flash (OpenCode and mini-swe-agent 61.8 % against OpenClaw 35.8 %), by 16.8 points on DeepSeek-V4.1-Flash (OpenCode and DeepSeek Harness 70.2 % against OpenClaw 53.5 %), by 16.0 points on Flint and by 35.4 points on Ripple-65B-Preview (Codex and Kimi Code 55.7 % against OpenClaw 20.4 %). The spreads are calculated from the unrounded values, so they can differ in the last digit from the difference of the numbers shown. “Up to 35 points” is therefore today's maximum, and it comes from Ripple. The cost column is the dollar amount per solved task across all four models, as the board reports it; Pi costs $0.76, OpenClaw $4.46, a factor of 5.9. The Ripple runs push the value up ($2.18 to $10.33 per solved task, depending on the harness). On GLM-5.3-Flash alone, one solved task costs $0.16 with Pi and $1.42 with OpenClaw, a factor of 8.9. On 11 November 2026 the largest spread was 37.4 points on DeepSeek; since then Hermes Agent and OpenClaw have been rerun, and the values of 11 November can no longer be reproduced from today's payload. The step that many sell as a model change sits one level lower, with the caveat that the operator could not separate the retrofit gain from noise.

Where the gap comes from

The largest single item is bookkeeping. Does the harness send the thinking sentences of the last answer back along again? OpenCode 1.19.0 spoke the Responses format to the built-in OpenAI provider, so the board had to bridge requests into Chat Completions. That bridge never produces the encrypted_content field the reasoning provider asks for. In the board log, the prompt of OpenCode 1.19.0 grew for 91 to 94 percent of call pairs whose answer came back with more than 300 tokens of thinking, and it grew by less than exactly that answer. For Codex, Pi, Kimi Code, Qwen Code, mini-swe-agent, Hermes and OpenCode 2 that rate sits between 0 and 1 percent. The article also states its limit: the operator writes that the point gain after the retrofit could not yet be told apart from normal variance.

Our measurement on our own gateway, 11 November 2026

We ran four harnesses through our own gateway against a vLLM instance on four cards (vLLM 0.32.0, TP4). The route answers with a reasoning_content field. Pi 1.5.2 and OpenCode 2.2.5 returned the thinking work in the second call, Pi with 12,103 characters of reasoning_content, exactly the amount from the first answer, OpenCode with 4,553 characters in the assistant block. OMP 18.18.7 and Claude Code 2.1.322 did not. OMP's assistant block was limited to role, content and tool_calls, reproduced twice. A text of about 250 tokens sent along inside reasoning_content lifts prompt_tokens from 70 to 334. The same text in reasoning_details lifts it to 87, so it is ignored there.

Assessment: The largest single improvement the board documents for one harness came without a new model. The operator traces it to a bridge that never passed thinking content on, and at the same time concedes that the retrofit gain could not yet be told apart from ordinary variance; that is no proof of the cause. For us that means: look first, then switch. Still open on our side is whether returning or discarding wins measurably on our own models and twenty multi-step tasks. A one-run comparison of two coding agents is noise, so two or three runs and then the mean. The board moves daily, so any figure only counts with its date: 15 November 2026. The tables, the noise calculation over 19.7 to 24.3 percent turnover per repeat (state of 15 November) and the cross-check from Surge AI are in the harness comparison post.

Deep Dive: OpenVDDK and getting VMware disks readable again

VMware brand mark
Broadcom removed the public VDDK downloads at the end of August 2026. OpenVDDK offers the library as a free reimplementation. (Quelle: Simple Icons 16.36.0, file vmware.svg (CC0), brand: VMware/Broadcom)

At the end of August 2026 Broadcom took down the public download pages for VMware's Virtual Disk Development Kit, without notice and without a replacement path. ShapeBlue documented on 25 August 2026 that every VDDK path variant in the portal returned only an error page, checked from Europe and through exit nodes in the United States, Brazil, Finland and Japan. Since then the instructions of virt-v2v, nbdkit and Apache CloudStack's VMware import end on a dead link. The developer rewrote the library in response. OpenVDDK rebuilds libvixDiskLib.so.8.0.3 independently, with the same functions and symbol versions, which is why tools can swap the file without adjustment.

What the repository documents

As of 11 November 2026: a LICENSE with the full Apache 2.0 text (11,358 bytes) and SPDX-License-Identifier: Apache-2.0 in every source file. A decision record (D-0027, accepted 5 October 2026) keeps Write, WriteAsync and WriteMetadata permanently out of scope; on a valid handle they answer with error 11. The compatibility matrix (state of 29 October 2026, vSphere 9.1.1 added on 2 November 2026) names vSphere 6.7 U3, 7.0.3, 8.0.3 and 9.1.1, each with vCenter and with direct ESXi sign-in, tested on AlmaLinux 9.8, Debian 13 and Ubuntu 24.04. All 43 public functions of VDDK 8.0.3 are documented one by one with status, including the ones the project has not implemented yet. Not tested: vSAN, vVols and Fibre Channel; vSphere 6.5 and older are excluded. State of releases: one tag v0.7.4 from 29 October 2026, 13 commits on main, two contributors, three issues, 58 stars (GitHub API, retrieved 11 November 2026).

The open defect on ESXi 7.0 GA

On 7 November 2026 an operator opened an issue with a typical migration setup: nbdkit with the VDDK plugin on a Proxmox VE node, pulling a VM's disk directly over NBDSSL from an ESXi host address, with no vCenter. That node runs ESXi 7.0.0 GA, build 15843807. Every attempt fails with bad frame magic 0x4. The maintainer confirmed the cause on 8 November 2026: 6.7 answers with NFC version 2, 7.0.3 with 9, 8.0 with 11 and 9.1.1 with 12. The GA release is missing from that list and, per the reporter's capture, answers with 5, while 0.7.4 already sends features and a resource pool query from version 3. The fix is announced for 0.7.5 and raises the threshold to 9; no 0.7.5 tag existed on 15 November 2026 (the newest tag is v0.7.4). The author calls 7.0 GA verified only after a run on a real GA node, and expects the builds between GA and Update 3 to behave the same way without having tested them. That is the second open point next to this one: OpenVDDK mirrors VDDK 8.0.3, while the VDDK 9.0 interface stays open.

Assessment: On the record, the reimplementation carries migrations because the source is only read and the target is written with its own code. The record is thin (two contributors, 58 stars, one open defect on 7.0 GA), and we have not tried the reimplementation ourselves; "carries" here means plausible per repository and project page, not confirmed by us. It does not carry a backup chain that replays through the same library. Anyone touching 7.0 nodes without vCenter should establish before the first run whether it is GA or Update 3, and plan for version 0.7.5 once it ships. The read rate of 2,488 MiB/s against VMware's 1,441 MiB/s comes from the project page, with its setup (HotAdd through vCenter 8.0.3, a worker VM with 3 vCPUs and 4 GiB, NFS 4.1 with RAM backing over 100 GbE, one request in flight, measured 3 October 2026). We did not measure it, so it stays a vendor figure with a scale attached. The OpenVDDK post holds the details, the compatibility matrix and the issue histories.

Deep Dive: EPYC 9006 Venice and the question behind the 1.6 terabytes

AMD brand mark
AMD publishes prices and bandwidth figures for EPYC 9006. We measured nothing ourselves; the numbers come from AMD and from the reports of Phoronix and ServeTheHome. (Quelle: Simple Icons 16.36.0, file amd.svg (CC0), brand: Advanced Micro Devices)

AMD announced general availability of EPYC 9006 "Venice" on 10 November 2026 as SP7 processors, the first server CPUs on Zen 6, built on TSMC's 2 nm process. The platform brings 16 memory channels per socket (RDIMM up to 8000 MT/s or MRDIMM up to 12800 MT/s), up to 8 TB of memory and 96 PCIe 6.0 lanes. The top model 9996 lists 256 cores and 600 watts TDP and costs $14,904 in the 1000-unit list price. Phoronix and ServeTheHome published their own reports with measurements the same day, and Level1Techs a video that we did not evaluate. We own no Venice hardware, so every figure here comes from those sources.

The bandwidth number and what it costs

AMD's 9686F datasheet lists 1,638 GB/s per socket, rounded to "up to 1.6 TB/s"; we did not find the advertising line itself in the AMD newsroom text or in Phoronix, only the datasheet value. That number applies with MRDIMM 12800 across all 16 channels and is a theoretical peak. The 9686F datasheet lists 1,024 GB/s with RDIMM 8000 and 1,638 GB/s with MRDIMM 12800, in both cases per socket at full population. MRDIMM is therefore optional, but the marketing runs on the value only the expensive population reaches. None of the sources we evaluated carried street prices for MRDIMM modules on launch day. Two sockets with MRDIMM 12800 would reach roughly 3.3 terabytes per second on paper per the 9686F datasheet (2 x 1,638 GB/s). That is again a datasheet peak; we did not evaluate measured bandwidth figures for this edition.

The rest of the package

The comparisons against the predecessor Turin, double throughput on integer work, 2.4 times on floating point, 2.7 times on memory bandwidth, come from AMD's briefings as relayed by Phoronix. They are vendor figures against AMD's own baseline. AMD's memory tiering whitepaper (MEXT) turns the calculation around: pages that have gone cold move to NVMe, and predicted ones come back before the application asks. It is measured with Synopsys PrimeTime, a signoff workload from chip design. The half population at 768 GB plus flash reached about 87 percent of the performance of a full 1.5 TB of DRAM, which the whitepaper calls 1.74 times the performance per DRAM dollar. That whitepaper measured the predecessor EPYC 9755 in October 2025.

Assessment: This platform is a memory offer, not a compute offer. Anyone serving local models with large contexts buys bandwidth and channels first. For that calculation the RDIMM population at 1,024 GB/s is the reference point; the advertised 1.6 terabytes needs MRDIMM. The smaller SP8 socket follows, per AMD, in the first half of 2027, with no known lane or channel counts. Nothing changes in our operation today: the two inference nodes we can measure run an EPYC 7502 and an EPYC 7F52, both SP3 with DDR4 (state of 15 November 2026, lscpu over SSH on two hosts). Venice is SP7, so a swap on the existing board is ruled out, and without a street price for MRDIMM no procurement calculation is possible. That leaves exactly one derived statement: the 1.6 terabytes is a number for the population list, not for the quote a reseller writes. The post with core data and an overview of the sources is the EPYC 9006 article.

Digest: other news from the week

Security

  • CISA KEV: 12 entries in one week The catalog of actively exploited flaws grew by 12 CVEs between 9 and 13 November, spread over entries from 9 November (four), 10 November (six) and 13 November (two). Counted on 15 November 2026 in the catalog's JSON feed (field dateAdded, feed state 13 November 2026, 1,781 entries in total). | reported (CISA catalog, recounted by us)
  • OpenSSH 10.6p2 closes the SSHole bypass, only Linux with musl and without PAM OpenSSH before 10.6p2 accepted any password between 256 and 1,024 characters on Linux with musl libc, without PAM, for accounts with a password hash set. The fix landed on 5 November 2026 and CVE-2026-123158 has been in the KEV since 9 November 2026. Measured over SSH on our Alpine host on 15 November 2026: it runs OpenSSH 10.2p1 with musl 1.2.5 and #UsePAM no, root has a hash set, and the package repository offers nothing newer. The flaw does not reach that box because PasswordAuthentication no is set, so the password path is never offered. | verified (release notes and NVD entry read, KEV entry in the JSON feed) and measured ourselves
  • Barracuda: two flaws make root on the mail gateway CVE-2026-116606 (authorization bypass in the configuration import, Email Security Gateway before firmware 9.4.0.030) and CVE-2026-111560 (command injection through the timezone field in six Barracuda products) combine into root on the appliance, but need the attacker to reach the administration interface (Barracuda: no login required, access to the admin UI is). Barracuda reports exploitation against a small number of organizations since the 9 November 2026 update, and NVD scores the first at CVSS 9.8 (Barracuda itself 9.3) and the second at 7.2 (Barracuda 8.6). The vendor advisory names patch IDs instead of versions (BNSF-40425 from 9.4.017, BNSF-40430 below it), a second notation for the same thing. | verified (vendor page and NVD entries read, 15 November 2026)
  • Microsoft's November update: about 1,100 entries on patch day, Tenable counts 1,093 In the MSRC feed (CVRF), 1,096 entries carry the first-published date of 10 November 2026; Tenable counts 1,093 CVEs on patch day, 63 of them Critical (1,029 Important). The feed of 15 November 2026 holds 1,644 November CVEs in total, because Chromium entries for Edge and further third-party components keep being added afterwards. The difference between 1,096 and 1,093 comes from the different scope of the MSRC feed and Tenable's count; we did not break it down. CVE-2026-124194, a spoofing flaw in Outlook, is exploited actively (CISA KEV since 10 November 2026; Microsoft and Tenable score it CVSS 6.5, Tenable rates it Moderate). | verified (Tenable read, count in the MSRC feed re-done by us on 15 November 2026)
  • Chrome 157.0.8103.50 closes an exploited use-after-free The stable update of 9 November 2026 brings 157.0.8103.50/.51 for Windows and macOS and 157.0.8103.50 for Linux, lists eight security fixes (six of them High) and closes CVE-2026-125350 (use after free in XML, CVSS 8.8), which Google says it knows an in-the-wild exploit for. The flaw has been in the KEV since 10 November 2026. | verified (Google blog, NVD entry and KEV feed read, 15 November 2026)
  • Node.js 24.22.3, 22.23.5 and 26.13.1: 16 CVEs, seven of them High The project published security releases for the 22, 24 and 26 lines on 11 November 2026 (HTTP/2, TLS and dependencies). Measured on 15 November 2026 across 48 of 49 hosts (node -v, docker exec): 61 of 66 containers with a readable Node sit below the fixed version (47 with an image name without a registry path, 14 with one; that is a rule of thumb for own versus third-party, not proof of origin), plus 10 of 27 hosts with Node in PATH; we could not read 185 of 251 containers. Derived, not proven: in the 66 readable containers the High cases have no entry point in our own apps, because Caddy terminates TLS and Node behind it speaks plain HTTP; a text search for HTTPS and HTTP/2 servers found five hits, none in an application of ours. That statement does not cover the 185 unreadable containers or the Node installations on the hosts. Tables and method: Node.js 24.22.3, 22.23.5 and 26.13.1: 16 CVEs and where our fleet stands. | vendor figure (announcement read), measured by us (fleet, 15 November 2026), exposure derived
  • ImageTragick2: image processing with the default policy reads and writes files CERT/CC note VU#580079 of 12 November 2026 describes two variants, MSL injection through SVG and a Ghostscript call through content detection, that give an application using ImageMagick without a restrictive policy.xml file read and write access in the process context; on the Ghostscript path access reaches only as far as /tmp in the observed setup, and code execution is possible only in some cases per the note. The note names Ruby on Rails, WordPress, MediaWiki, ownCloud Classic, Redmine and pipelines with image pre-processing as affected. Our write-up with measurement: ImageTragick2: A policy.xml decides whether ImageMagick reads your files. | reported (CERT/CC note read, 15 November 2026)

Windows and Microsoft

  • Microsoft 365 went down worldwide on 12 November: Exchange Online, Teams, Copilot, Defender XDR Per Microsoft's history notices, the Exchange incident EX1550543 ran, per its closing text, from 12 November 2026, 16:20 UTC, to 13 November, 16:10 UTC, about 24 hours; the header of the companion incident MO1550821 gives 12 November, 10:05 UTC as its start, which would make it about 30 hours. The companion incident also hit Teams, SharePoint Online, Copilot, Defender, Purview and Intune. As a preliminary cause Microsoft names a latent fault in the systems that accept connections, triggered by traffic shifts during network maintenance; it rules out an attack, and the post-incident report was still pending when we read it. Microsoft extended the mail timeout to 48 hours. Here is what reached us (counted read-only on 15 November 2026): the Zammad inbox reading through Microsoft Graph logged 1,634 retrieval errors from 12 November, 17:42 UTC, to 13 November, 13:39 UTC, and the mail gateway deferred 3,198 deliveries to our tenant (87 queue IDs); all 87 were delivered later, and no bounce came from the outage. Timeline, method and checklist: The Exchange Online outage of 12 and 13 November: timeline, Microsoft's stated cause and what reached us. | reported (Microsoft's history notices read via the University of Pennsylvania mirror and BleepingComputer, our own admin center not) and measured by us (Zammad host and mail gateway, 15 November 2026)
  • Windows 11: the November update KB5132061 fails on some devices with 0x800f0989 On 12 November 2026 Microsoft confirmed that the November security update (KB5132061) fails on Windows 11 devices that installed the October preview update KB5127936, with errors 0x800f0989, 0x800f0993 or 0x80070306, and the installation can lock up midway. Microsoft names an automatic repair within 24 to 48 hours from 11 November 2026, 10 am PT, after which the update installs by itself; anyone still seeing errors after 48 hours should contact support. Patch day itself is in the security section above, PowerShell 7.7.0 below under releases. | reported (BleepingComputer of 12 November 2026 read, Microsoft's notice not)

Hardware

  • ServeTheHome on EPYC 9006: up to 256 cores, 16 memory channels, PCIe 6.0 ServeTheHome (10 November 2026) shows the final SP7 data on launch day; the price ($14,904), availability and the SP8 date (first half of 2027) are confirmed by Phoronix and AMD's newsroom. The deep dive above holds the bandwidth arithmetic and our reading. | verified (ServeTheHome, Phoronix and AMD pages read, 15 November 2026)
  • Google extends the TPU 8 line: 8i Max with two chips per package The 8i Max puts two 8i chips on one package and doubles compute, memory and bandwidth per package, and a Confidential TPU 8tx with 5th-generation EPYC (Turin) hosts joins the line. The step from 384 MB to 768 MB of SRAM is StorageReview's calculation, and Google does not name the package's HBM size. | reported (StorageReview of 13 November 2026 read, Google's own page not)
  • Nvidia's Vera CPU in its own measurement: 2,073 GB/s in STREAM triad StorageReview measures 2,073 GB/s across the two-socket MGX reference system with SOCAMM2 memory, which is 86 percent of the 2.4 TB/s specified for two sockets; one socket alone reaches 1,055 GB/s, 88 percent of the 1.2 TB/s specified per socket. The comparison against five x86 servers, including the build defect in that comparison (bare GCC without streaming stores, which drags the x86 candidates down), is in the post NVIDIA Vera CPU: what an independent lab measured. | measurement by StorageReview, 10 November 2026
  • Lenovo puts 96 EPYC 9006 sockets into one liquid-cooled rack The ThinkSystem SC755 V5 brings 48 dual-socket nodes, up to 24,000 cores and 192 TB of DDR5 on two floor tiles, spread over three 13U enclosures with direct warm-water cooling. That is a configuration for data centers with cooling built in, not for the closet next to the office. | vendor figure (StorageReview of 11 November 2026 read)

AI and development

  • llama.cpp ships v1.0.0: API cleaned up, FP8 KV cache, hardening behind flags The first 1.0 on 13 November 2026 removes 35 deprecated functions from the public API, moves speculative decoding (NextN) into llama.h and adds KV caches in the OCP FP8 format. In the server come hardenings against CORS misconfiguration, DNS rebinding and SSRF through the CORS proxy and media URLs; they change behaviour only when --cors-origin is set or for tools, the maintainers themselves see no vulnerability in them, and how 1.0.0 answers with the default settings is untested. Our standby lane runs the container image ghcr.io/ggml-org/llama.cpp:server-cuda at build state 11 November 2026, so before this release; the update is not applied here (15 November 2026, docker inspect on the reserve host). The build is b12282 of 11 November 2026, 03:31 UTC; the four server-hardening PRs were merged only afterwards (11 November, 20:17 UTC to 12 November, 11:33 UTC) and are not in it. Measured and put in context: llama.cpp 1.0.0: three ZDI flaws on release day and a server with open defaults. | vendor figure (release page read), container state measured by us
  • Three ZDI reports on code execution in GGUF parsing, on the day of 1.0 ZDI published three llama.cpp advisories on 13 November 2026: ZDI-26-976 for a heap buffer overflow in GGUF parsing (CVE-2026-126975, CVSS 7.8), ZDI-26-975 for an integer overflow in the same parser (CVE-2026-126974, 7.8) and ZDI-26-954 for an integer overflow in the mtmd multimodal path (CVE-2026-126976, 7.0 with a higher attack complexity). All three need user interaction: someone has to open the file. The two 7.8 notices name pull request 31336 as the correction, the multimodal notice points to commit e2e7a9b2, and the v1.0.0 release text of 13 November 2026 lists none of the three CVEs. On our side the git history says the fixes are already in the build: pull request 31336 (merged 28 October 2026) and commit e2e7a9b are both ancestors of b12282. That is derived from the commit history and no exploit test; we did not check whether the code covers every single ZDI description. Anyone running llama-server against unknown files should check their own build and its change date. | reported (three ZDI advisory pages and the release text read, 15 November 2026; the CVE records at CVE.org and NVD were not yet retrievable on 15 November 2026), derived for our build (git history)
  • Jensen Huang resigns as Nvidia CEO, Colette Kress takes over on an interim basis Per Nvidia's statement of Thursday 12 November 2026 (as relayed by The Decoder), Huang resigned with immediate effect as CEO, president and board member; Colette Kress, until then CFO, is interim CEO. Since the end of October roughly $1.4 trillion of market value has gone, per The Decoder; the ARD background report on tagesschau.de of 12 November 2026 describes the same situation. The reports name allegations as the reason, which we do not repeat here: we have read neither Nvidia's statement nor a statement from the people concerned. | reported (The Decoder and tagesschau.de read, 15 November 2026; Nvidia's statement not)
  • Inkling 2 Small: open mixture-of-experts model with 594 billion parameters and a 1 million token context Thinking Machines publishes the model on Hugging Face under apache-2.0. Measured through the API on 15 November 2026: 25 safetensors files with 393,510,049,944 bytes (366.5 GiB), 594,050,672,720 parameters, model_max_length 1,048,576 tokens. For scale: 366.5 GiB of weights is eleven times a 32 GiB card, so this is no candidate for us, only the measure of the order of magnitude. Our fit test against the fleet is in Inkling 2 Small: 366 GiB of weights against our 127 GiB of VRAM. | measured (file sizes and parameter count from the Hugging Face API, 15 November 2026)
  • Meta serves Muse Ember at $0.10 per million input tokens Up to 200,000 tokens of prompt the hosted model costs $0.10 per million input and $0.50 per million output tokens, above that $0.30 and $1.50, and the provider states a context up to 1 million tokens. | reported (Simon Willison of 10 November 2026 read, Meta's own page not)
  • Amazon shows the arithmetic: a decision model instead of Claude Sonnet in ticketing, 44 times cheaper The post sets $22,200 per million tickets at Sonnet 4.6 prices against roughly $500 per million tickets with Strands Decider, at 640 milliseconds per request and about 5,600 tickets per instance and hour. That is a vendor calculation on their own tasks, without an independent measurement. | reported (15 November 2026)

Consumer hardware

  • BSI warns of critical flaws in the Linksys EA8100 router Advisory WID-SEC-2026-4555 (published 12 November 2026 at 23:00 UTC, so overnight into 13 November in Germany) rates five flaws (CVE-2026-127273 to -127277) as critical: a remote, anonymous attacker can bypass safeguards, read information and manipulate data. One model is listed as affected, the EA8100 with firmware 1.0.3.216940. The BSI document does not say whether a patch exists. For a device in a customer network that means: check the model and firmware level. We read the BSI's CSAF document. | reported (BSI CSAF document read, 15 November 2026)
  • Synology C2 Identity: patched since December 2025, ZDI and Synology score it differently ZDI published the Pwn2Own finding CVE-2025-14713 (network-adjacent, no authentication, DiskStation DS925+) on 13 November 2026 at 6.5, while Synology itself has scored the same flaw at 7.5 with a network vector since its notice of 16 December 2025 (last updated 27 May 2026) and delivers the fix from package level 1.76.0-0307 for DSM 7.1, 7.2.1, 7.2.2 and 7.3; the advisory's product table says “Critical” and its detail section “Important”, without explaining it. What is new is the ZDI publication, not the patch. Both figures stay side by side, because both come from the parties responsible. Our write-up: Synology C2 Identity Edge Server: CVE-2025-14713 and why two CVSS scores don't contradict each other. | verified (both primary sources read, 15 November 2026)
  • The FCC hands out one-year exceptions to the router ban, TP-Link is left out Per heise (14 November 2026) the FCC approved an application by the cable association NCTA covering changes to already-approved devices that would otherwise lose their grandfathered status; the exception runs one year for the association's members and suppliers. Vantiva may sell its fiber router without a barometer and GPS, and TP-Link has not received a single exception despite announcing $1 billion of US manufacturing. | reported (15 November 2026)
  • Unraid OS 7.4 gives Docker its own networks and the Mover a progress bar Version 7.4 of 11 November 2026 lets containers join several custom networks through their template and adds progress and status tracking for the Mover; it also brings a memory-limit field for containers, preserved hard links in the Mover and kernel 6.18.60. | reported (linuxiac read, Unraid's own announcement not)
  • Pi-Dock-Wifi: a home router from a Pi 5, two wifi cards and Pi-hole The open project bundles hostapd, dnsmasq, firewall and routing in containers on a Raspberry Pi 5 with two radio modules (2.4 and 5 GHz) and ships network telemetry through Pi-hole and a ttyd terminal. The author says the full installation instructions have not yet been independently validated on a fresh system; the source gives no throughput figures. | reported (noted.lol read, 15 November 2026)

Software releases

  • PostgreSQL: a maintenance wave on the day version 14 ends PostgreSQL 18.7, 17.12, 16.16, 15.20 and 14.25 arrived on 12 November 2026, with 52 security fixes across all lines (40 of them in version 14, 29 with CVSS 8.0 or higher) and more than 140 further corrections, plus 19 RC 2 the same day and version 19 final announced for 19 November 2026. The 14.25 is the last release of the 14 line ever; after it, 14 gets no further patches. Our fleet check: PostgreSQL 14 has reached end of life: a fleet check across 14 measured databases. | vendor figure (project announcement of 12 November 2026 read)
  • Stalwart 1.1.0 adds passkeys and one upgrade trap The mail server 1.1.0 of 9 November 2026 adds passkeys (WebAuthn) for accounts in the internal directory. Anyone upgrading from an older version has to add the new sysPasskey permissions, either by hand through the recovery administrator or with a one-time restart under STALWART_UPDATE_ROLE_PERMISSIONS=true (then remove the variable), or the default roles will not have them. We run version v1.0.2 in the container (measured 15 November 2026 on the mail host), the upgrade is not applied. Our write-up with measurement: Stalwart 1.1.0: passkeys, and the role permissions an upgrade leaves out. | vendor figure (9 November 2026)
  • Nextcloud Hub 26: maintenance releases for all three branches Nextcloud published levels 33.0.11, 34.0.6 and 35.0.3 for Hub 26 on 13 November 2026, with error corrections and stability and security improvements. | vendor figure (13 November 2026)
  • Nginx Proxy Manager 2.17.0 with security and SSL fixes Version 2.17.0 came on 12 November 2026 with updated Certbot DNS plugins (Hetzner, Infomaniak and Akamai Edge DNS among them) and a new Scaleway plugin, and the release notes warn about breaking changes in the DNS plugins for EdgeDNS, Beget and Infomaniak. Our internal NPM lane runs npm-node:2.16.0-n22.23.4 (measured 15 November 2026 with docker ps), the update is not applied. | vendor figure (GitHub release read, 12 November 2026; linuxiac reports the same), our own version measured by us
  • Immich 3.4.0 adds JPEG XL, 3.4.1 follows the next day Immich 3.4.0 of 10 November 2026 adds JPEG XL, workflow improvements and several security fixes; patch 3.4.1 of 11 November 2026 fixes among other things the performance problems with shared links. | vendor figure (both GitHub releases read, 10 November 2026)
  • PowerShell 7.7.0 seals off the constrained language mode PowerShell 7.7.0 arrived on 10 November 2026, secures bypasses in the constrained language mode, hardens remoting and now warns when a file catalog is signed with SHA-1 (build basis .NET SDK 11.0.100). | vendor figure (GitHub release read, 10 November 2026; windowspro.de reports the same)

Smart home

The house control sources brought two solid items this week. Nothing with substance turned up in the sources we read about Loxone, Matter or Thread, so it stays at two.

  • Pwn2Own findings on the Philips Hue Bridge: CVSS 8.8 for code execution in the LAN ZDI published four advisories for the Bridge on 13 November 2026, the highest at 8.8 (integer underflow in chacha20_decrypt of the HomeKit accessory protocol; a buffer overflow in pairing is also scored 8.8, plus 8.1 for an authentication bypass and 6.5 for an information disclosure). All of them are usable from the local network without authentication, which is exactly the network most Bridges sit in. ZDI names Bridge v2 software 1975170000 as the fixed level; the Hue release notes list that level since 16 February 2026 without a security note, and the latest level is 6.5.1978688000 of 12 November 2026. The Hue app shows which version your own Bridge runs; check that automatic updates are on. Our write-up: Philips Hue Bridge: Four Pwn2Own Flaws in the HomeKit Service, Check Your Firmware. | reported (all four ZDI advisory pages read, no second newsroom)
  • c't lets a language model run Home Assistant administration The hands-on article builds dashboards and complex automations by prompt over the unofficial MCP interface and names prompt injection as the central risk, with local models tested as an alternative lane. That is the same boundary we drew in Prompt injection: limiting agent rights. | reported (15 November 2026)

From the blog

More from the week

Sources for this edition

Compiled on Monday, 16 November 2026. Sources: CISA, NVD, OpenSSH, Barracuda, Tenable, Microsoft Security Response Center, BleepingComputer, Google Chrome Releases, CERT/CC, AMD, Phoronix, Tom's Hardware, Level1Techs, ServeTheHome, StorageReview, GitHub (llama.cpp, Stalwart, OpenVDDK, Immich, Nginx Proxy Manager, PowerShell), Zero Day Initiative, The Decoder, tagesschau.de, Hugging Face, Simon Willison, Strands Agents, BSI/CERT-Bund, Synology, heise online, linuxiac, noted.lol, PostgreSQL.org, Nextcloud, windowspro.de, ShapeBlue, The Register, Surge AI, Harness Leaderboard, Terminal-Bench, DeepSWE, Simple Icons. From the week's posts come in addition: Microsoft Learn, Azure status page, University of Pennsylvania (mirror of Microsoft's notices), Postfix, SecurityWeek, Node.js, Docker Hub, Mitre CVE, FIRST, TrendForce, Geizhals, Notebookcheck, DeepSeek API Docs, Pandaily, Home Assistant, ESPHome, YouTube, PyPI, Hacker News, arXiv, W3C WICG, Linux Foundation, gyptazy, Proxmox forum, Facebook/Meta, Il Fatto Quotidiano, Garante per la protezione dei dati personali, CBS News, TechCrunch, Google and Apple support, Android Developers, Wikimedia Commons, Philips Hue, Stalwart, ImageMagick, NVIDIA Developer, CRIU, vLLM, SGLang, Thinking Machines, Artificial Analysis, Anthropic, RunPod, Azure Retail Prices, AWS Pricing, Bambu Lab, CNC Kitchen, Prusa, EUR-Lex, Curia (CJEU), ETSI, TechSpot, ITDaily, dig.watch and Apache. Anything that arrives after that goes into the next edition.