IT-News Week 48: R9700, Timestral, vLLM 0.34.0
This week is about numbers you can recount. First, our dual-R9700 box: Radiance 2.1.0 runs Qwen3.8-27B in tensor-parallel mode at 89 to 104 tokens a second on a single stream, four to five times what vLLM on ROCm reached here, and the deep dive says why that comparison still needs a steady hand. Second, Timestral: Mistral reported first place for its time-series models, we recalculated the scores ourselves, and the values land to the last decimal while the leaderboards say something else. Third, vLLM 0.34.0: the new default allocation of the KV cache changes what a gpu-memory-utilization figure means at all, and one switch we set long ago leaves our lane alone. Label key: verified (two independent sources read), reported (one source read, even where we calculated figures from it), vendor figure, derived (our reading). Every entry names the source we actually opened.
On the week's output: we published 14 posts. Three of them are the deep dives above, the other eleven sit under "From the blog". The version and release states in the digest were measured again on 22 November 2026, and every entry says which source we opened. The digest is news coverage; where an entry says "checked by us" or "measured by us", it is our own measurement.
Deep Dive: Two Radeon R9700 on Radiance and what the measurement really shows
On 17 November 2026 we published the post about our AMD lane: two Radeon AI PRO R9700, Qwen3.8-27B, four engine setups head to head, measured on 16 November. The shortest version of it: Radiance 2.1.0 leads in every mode. On one card that is 59 to 61 tokens a second on a single stream, against 19 to 28 for vLLM on ROCm 7.2 and 38 for llama.cpp with MTP speculation. Run as one tensor-parallel instance over both cards, the single stream sits at 89 to 104 tokens.
| Engine (state of 16 November 2026) | Single stream, tok/s | 8 parallel, tok/s | 16 parallel, tok/s |
|---|---|---|---|
| vLLM v0.33.0 on ROCm, 1 card | 19 to 28 | 69 | not measured |
| llama.cpp with MTP, 2 cards | 38 | 109 | not measured |
| Radiance 2.1.0, 1 card | 59 to 61 | 266 | not measured |
| Radiance 2.1.0, 2 replicas on 2 cards | 57 (at 2 streams) | 371 | 546 |
| Radiance 2.1.0, tensor-parallel over 2 cards | 89 to 104 | 405 | 363 |
The catch in that table: Radiance ran with a drafter (speculative decoding) and a different quantization than vLLM, 4.9 bit against AWQ-INT4. We did not separate the causes, so this is no clean engine comparison. The direction on two cards matters just as much: at 8 streams tensor parallel leads (405 against 371), at 16 streams the two replicas turn it around again with 546 against 363. Long answers for one agent are served better tensor-parallel; many short requests at once are served better by two replicas.
Tensor parallel also has the memory argument. The weights split across the cards and each one keeps 19.46 GiB for the KV cache. The pool holds 1,103,104 tokens, which is 4.2 sessions of 262,144 tokens of context each. To put that in scale: Qwen3.8-27B carries token KV in only 16 of its 64 layers, and at 8 bit that is 32,768 bytes per token (a calculation from the model configuration, we never measured it). Radiance 2.1.0 knows Int8 and FP8 as KV formats and nothing down to 4-bit KV; the MXFP4 result shown in the post (4.4 sessions) runs on Radiance 1.17.0, because the 2.x line refuses MXFP4 containers. Starting two of these 27-billion models at once inside one tensor-parallel pool fails at about 29 GiB of use per model.
Assessment: In our measurement the AMD lane moved from a footnote to a usable second path, and that is the actual news. What it does not deliver: proof that AMD computes faster than CUDA. Quantization, drafter and kernel set were all different, and the load test of our production traffic over the new path (of 12 parallel requests, 2 ended in an error because both tiers were full) shows work left before a full production run. The single stream is bandwidth-limited, the aggregate figures hang on parallelism, and anyone recounting has to name the column. Every number here dates from 16 and 17 November 2026, our own measurement. The post with the protocol, the conversion times and all boundary conditions: Two Radeon AI PRO R9700 with Radiance.
Deep Dive: Mistral's Timestral: the numbers hold, the first place does not
Mistral introduced Timestral on 15 October 2026, two small time-series models (209M and 123M parameters), and reported first place across all model categories on GIFT-Eval. We went and recounted. The reported values 0.658 (MASE) and 0.446 (WQL) show up in the board's result files as 0.6575 and 0.4460, and aggregating the individual runs reproduces them to the fourth decimal. The fev figure 50.4 is covered too: the fev table carries 50.39, and there Timestral-209M sits at place 4 of 43, behind TimeRSI-1.5 (51.82), TimeRSI (51.24) and TimeH2O-v1 (50.83). On the GIFT board the model is 14th of all 163 model directories; among the open zero-shot models it is second: TimeH2O-v1 leads at 0.6536 gMASE, Timestral-209M follows at 0.6575, with t0-gamma behind at 0.6608. The gap to t0-gamma behind it is 0.5 percent on MASE and 1.3 percent on WQL, the range where a different task mix flips the order. We read both boards on 23 October.
The licence question stayed open, and open because two primary sources contradict each other with neither explaining how: the repository says Apache 2.0, while the licence file itself carries a "Modified MIT" with a threshold of $20 million in revenue per month. We read both places on the same day and cannot resolve the difference. Then the state of the project: last commit on 15 October, 1,231 stars, 837.27 MB on Hugging Face at 30,653 downloads, no package on PyPI (an open issue asking for one has sat unanswered for weeks).
Assessment: Our rule for leaderboards applies here as well: a rank is a property of an interface on a given day, the number a property of the model. Anyone working with Timestral gets solid values, plus a project untouched for weeks, no package in any index and an unclear licence. For a pilot that means: settle the licence question with the vendor and put your own data through your own measurement instead of through a board ranking. The post with the calculation path, the state of the tables and all boundary conditions: Mistral's Timestral: the numbers hold, the first place does not.
Deep Dive: vLLM 0.34.0: the KV cache default changes
vLLM 0.34.0 landed on 16 November 2026 at 23:22 UTC and changes what configuration values mean, in one place fundamentally. First, the extensible KV cache is the default now: the pool grows into free memory in blocks of 64 tokens instead of being sized once at startup. Second, gpu_memory_utilization now stands at 1.0 by default. Both together move the default behavior of every installation that never set the value on purpose. Anyone wanting the old fixed pool sets --no-enable-extensible-kv-cache (vLLM then reports its utilization estimate the classic way again, 0.92 with us) or hands the engine a hard byte count through --kv-cache-memory-bytes; that switch wins and turns the extensible allocation off, exactly as config/cache.py spells it out in the source. A later state added a warning (PR 66704) for anyone mixing the two paths. Two further changes ride along in the same release: the delivery packages for CUDA are now built against CUDA 13, and the ReDoS fix #52120 is in v0.34, so anyone staying on v0.32 or v0.33 does not have the atomic group in the regex yet.
Our own lane (four cards, production) has been running for weeks on --kv-cache-memory-bytes 7600000000 as a hard ceiling. Measured on the day of the post: with that switch set, the update changes nothing in behavior, the allocation stays as it was, and the warning about mixing the switches does not reach us.
Assessment: For operators the lesson is first of all this: switches do not inherit, they act. Anyone who ever set gpu_memory_utilization below 1.0 to hold GPU memory back has to check after the update whether the assumption still holds; the defaults are the real breaking change. For anyone running fixed pools, as we do, the allocation change is a non-event. The ReDoS line is different, because it concerns every version below 0.34 that serves URL-driven tokenizer paths. Details, source locations and our measurement: vLLM 0.34.0 measured.
Digest: other news from the week
Security
- Google's threat hunters see AI attacks with no human in the loop The Google Threat Intelligence Group and Mandiant wrote on 16 November 2026 in the AI Threat Tracker that attackers had, "for the first time", fully automated vulnerability discovery and exploitation without any human hand in it. The report itself names no percentage for that; the figures Google and Mandiant do bring are a zero-day rate averaging 11 a month over the last three months (8 a month before), the PaperCut case where exploitation came four days after disclosure, and LLMJacking cases above $100,000 a day. | reported (Google and Mandiant post of 16 November 2026 read, 22 November 2026)
- FBI warning on ARTEX: an LLM pentest tool, under three hours to a takeover The FBI reported ARTEX on 13 November 2026 in a FLASH advisory as an open-source agentic pentest framework written in Go, with which attackers take over IT networks in under three hours; the timing comes from Microsoft (21 October 2026) and is quoted inside the FBI report. Reported alongside it: more than 1,600 openly reachable ARTEX consoles, and the path through ConnectWise Cloud Backup into over 2,250 tenants. heise online wrote on 20 November 2026 about days of outages after an attack on the European Commission with ARTEX suspected; neither the FBI nor heise names a Chinese APT as the party responsible. | reported (FBI PDF and heise article read, 22 November 2026)
- Java4Shell: CVE-2026-114938 in OpenJDK, patched the day of the warning Oracle's Critical Security Patch Update of 17 November 2026 warns of remote code execution in the JDK core library; affected is OpenJDK 17.0.6 through 27. Exploitation needs an x86-64 process with AVX-512-VBMI instructions and a 64-character pattern of A plus an
=in the data being processed; ARM64 is unaffected. Two ratings stand side by side because both come from parties entitled to give one: Oracle and NVD list CVSS 3.1 at 8.1 (High, "difficult to exploit"), the reporter Calif gives CVSS 4.0 and 9.2 (Calif post). CISA carries the flaw in an alert of 17 November 2026; in the catalog of actively exploited vulnerabilities it does not appear to this day. The crash had been open as a bug report since June, and the fix pull request has sat unmerged ever since. | verified (Oracle warning and reporter post read, 22 November 2026) - NGINX closes ten flaws, fixed in 1.30.7 and 1.31.9 The advisories of 18 November 2026 cover HTTP/2 handling and module paths; NGINX Plus moves from R34 through R36 up to 37.x, while the 1.28 and 1.26 lines are still open on support and stay without a fix. The highest rating belongs to CVE-2026-123007 (F5 gives 9.2 on CVSS 4.0 and 8.1 on 3.1): a buffer overflow in the HTTP/2 compressed-header handling, and code execution only where ASLR is switched off on the machine. A reload is not enough, swapping the binary takes a restart. Our own pre-check of the edge proxy came back: not one HTTP/2 listener active, so the condition for the worst case does not apply. | verified (NGINX advisories read, 22 November 2026), fleet state checked by us
- A KEV week with nine entries; the LiteLLM listing is seven months old The catalog of actively exploited vulnerabilities grew by nine CVEs between 17 and 20 November 2026, all without an attribution to an attacker and none from the ICS world. One entry stands out: CVE-2026-40217 (LiteLLM, CISA deadline 2 December 2026), where the first version that fixed it, 1.83.9, is roughly seven months old and the flaw needs a valid proxy admin key (CVSS 8.8, PR:L). The genuinely new security fix of the week is a notice of 20 November 2026 (moderate tier, a metadata read path through
/key/info, fixes in 1.106.12, 1.108.13, 1.109.3 and 1.110.2). Between 16 and 21 November 2026 there were 18 LiteLLM releases, 13 of them stable. | verified (KEV JSON feed and advisory read, 22 November 2026) - Project Zero shortens the disclosure window: 90 days now, 30 days from 2028 The policy updated on 13 November 2026 sets full disclosure after 90 days as today's standard (before: 90 days plus a 30-day grace period), 60 days from 2 February 2027 and 30 days from 3 January 2028; for actively exploited flaws it stays at 7 days. | verified (policy page read, 22 November 2026)
- Four days of outage on the NSA network, reportedly from a patch load Golem reported an outage of roughly four days on the classified NSAnet; heise reports the same timeline for US banks and names "too many patches" as the cause given. None of that is confirmed officially; as of 22 November 2026 there is no confirmed root cause. | reported (Golem and heise read, 22 November 2026; no confirmation from the NSA)
- Microsoft's agent-framework on PyPI: two versions with a credential stealer Versions 1.26.1 and 1.26.2 of 17 Nov 2026 ran malware through a
.pthfile on every Python start, collecting environment variables and credentials for Azure, GitHub Actions and other services; the last clean version is 1.26.0. According to the reports both versions were isolated after about an hour, and Microsoft made no statement in the sources we read. We do not use the package; it was not found in our repos or in ten checked containers on two Linux hosts (a subset, not full proof). | verified (Wiz, Aikido, StepSecurity, Socket, BleepingComputer and GitHub advisory GHSA-764m-3ppc-5m73 read, 22 Nov 2026), own fleet state checked
Windows and Microsoft
- Microsoft's preliminary report on the cloud outage: over 80 percent of gateways stopped answering The preliminary post-incident report (dated 15 November 2026) traces the outage of 12 and 13 November back to a latent fault from a July fix on the connection-accepting side; at the peak more than 80 percent of gateways stopped answering. The final report announced for 20 November 2026 was still missing when we looked on 22 November 2026. In parallel, Microsoft reported on 16 November 2026 a phishing wave that uses the outage as bait. | reported (Microsoft report through a mirror carrying the date and heise read; final report open)
- A cloud outage at an amusement park: 15 injured after a chlorine fault at Phantasialand According to WDR's research, the dosing trip caused by the Microsoft outage on 12 November 2026 led to an over-concentration of chlorine with 15 injured, 4 of them taken to hospital. Microsoft ruled out a cyber attack on the plant; what remained was a technical defect in the connection. | reported (WDR read, 22 November 2026)
- Windows Server 2028: the public preview is here Microsoft put the first public preview into the Windows Server Insider on 17 November 2026; the preview pages name new security building blocks and list the AMD MI350P as a GPU-throughput preview "from the first half of 2027", with no date for the finished product. | verified (Learn pages read, 22 November 2026)
- Windows 11 reports coding agents to Microsoft: what the telemetry experiment really sends The post of 17 November 2026 documents inside Windows' collection code (the
MissionControlAggregatorcomponent) that the agent interface of coding tools goes to Microsoft together with an image of the window: bitmaps of the CLI windows, no file contents and no keystrokes, and that it is a server-side controlled experiment. Microsoft had not answered by 22 November 2026; Golem reported the same day. | verified (source-code analysis read, cross-checked with Golem, 22 November 2026) - Remote Lock: the laptop gets locked from a distance The Insider builds of 16 November 2026 bring the preview of Remote Lock: an IT administrator can lock a lost or stolen device until it is picked up again. | reported (Windows Insider blog read, 22 November 2026)
- Ignite 2026: MAI models on Azure only, Cobalt 200 available Microsoft introduced licensed model weights on 17 November 2026: use through Azure, no public weights to download. At the same time the Azure pages report the new Cobalt 200 generation as generally available. | verified (microsoft.ai and the Azure blog read, 22 November 2026)
- Aftermath of the November patch: activation errors, PowerShell mail, a boot loop Three spots on the same update train: an out-of-band update of 13 November 2026 fixes the licensing error 0x80041008 (SLMGR). KB5131085 on Server 2019 disturbed mail sent through
Send-MailMessage, because NTLM signing is required now. And KB5132061 puts individual Windows 11 devices into a boot loop through thePxHlpa64.sysdriver. | reported (Borncity read, 22 November 2026) - SharePoint: CVE-2026-124223 allows code injection The MSRC entry of 17 November 2026 lists a flaw in SharePoint Server with CVSS 8.8 through which an authenticated attacker can inject and run code. | verified (MSRC entry read, 22 November 2026)
Hardware
- Nvidia shows the Rubin platform at SC26: over 50,000 CUDA cores, available 2028 Nvidia's post on the SC26 (Chicago, 15 to 20 November 2026) lists the HPC variant of Rubin with "more than 50,000 CUDA cores" and predicts availability "expected 2028" for it; the post names no memory size, floating-point performance or power draw. The much-quoted "3x" formula stays unusable without a baseline. | vendor figure (Nvidia post read, 22 November 2026)
- Vera CPUs for science: 192 CPUs per rack, systems from early 2027 For its science program Nvidia names 192 Vera CPUs per rack with Quantum-X800 InfiniBand, over 200 TB/s of aggregate memory bandwidth, 88 of its own Olympus cores per CPU, up to 1.2 TB/s of memory attach over SOCAMM2, roughly 300 TB of memory and 1 PB of storage per rack; systems from Dell, HPE, Lenovo and Supermicro "from early 2027". The 1.9 and 2.4 times figures for OpenFOAM come from Nvidia's own test, and for WRF Nvidia gives 1.6 and 2 as an expectation, a measurement it was not; do not mix the two. | vendor figure (Nvidia release read, 22 November 2026)
- Co-packaged optics at Lambda: over 10,000 GPUs in Quebec Lambda reports what it calls the "world's first" large-scale deployment of optics-attached switches with over 10,000 GB300 NVL72 accelerators in Quebec, Canada. The numbers around it: Nvidia names 4 percent and about 10 percent overhead, Lambda measures exactly 3.8 and 10.1 percent. On reliability, 15 times (the SC26 statement) and 10 times (August 2025) stand side by side as vendor figures. | vendor figure (Lambda and Nvidia posts read, 22 November 2026)
- Intel's SC26 appearance: the D270 GPU with 160 GB and Diamond Rapids on the stage slide The data center page lists the D270 datacenter GPU with 160 GB of LPDDR5X at 1.5 TB/s, "scalable to 480 GB", 350 watts, PCIe 5.0 and air cooling; the "up to 5.9 times" against Nvidia's RTX PRO 6000 (600 W) comes from "pre-production hardware, an Intel test". The stage slide at StorageReview gives Diamond Rapids 1.6 TB/s, 256 P-cores and 1.28 GB of last-level cache; a market entry date is missing, and no Intel announcement carries these numbers. Converting the figures against each other is our own derivation. | verified (Intel page and StorageReview read, 22 November 2026), ratio derived
AI and development
- GPT-6.2 Astra and Sol: OpenAI crosses its own "High" threshold for self-improvement OpenAI took two models live on 17 November 2026 and calls Astra "the first externally released model to reach the High threshold for AI self-improvement" (quoted in translation). Prices per million tokens per the provider: Astra $8, $0.40 and $40, Sol $2, $0.10 and $10 (input, cached input, output); the safety page says nothing about context size, so that stays unchecked on our side. | verified (OpenAI safety page read, 22 November 2026; Simon Willison's take cross-checked)
- Qwen4-Max: a big model, but no open weights, and no win over GPT-6 Alibaba presented Qwen4-Max on 16 November 2026: a backbone of 2.4 billion and 71 billion active parameters, 1 million tokens of context, $1.50 and $4.50 per million tokens. The report that the model beats GPT-6 does not survive the Artificial Analysis table: on index version v5.0.1 Qwen4-Max sits at 43 points, Claude Fable 5.2 at 65. Open weights do not exist. | verified (Qwen blog read, 22 November 2026)
- Hugging Face passes model routing on to OpenRouter On 16 November 2026 Hugging Face replaced its own model routing with the one from OpenRouter: billing through the Hugging Face key with no surcharge, but breaking for applications that set provider-specific keys, and Pro costs $2 a month. As side figures the post reports 46 percent fewer capacity errors, 16.9 percent shorter startup time and 16.3 percent more throughput. In the same week Nvidia announced it wants to acquire Hugging Face. | verified (HF post read, Nvidia announcement cross-checked, 22 November 2026)
- Nvidia's numbers: $115.8 billion in revenue, compute as nearly all of it The 8-K filing of 17 November 2026 lists $115.8 billion in revenue (up 20 percent quarter on quarter, up 103 percent year on year), $107.6 billion of it in the data center segment; that is 92.9 percent (our derivation). The outlook is $130.0 billion plus or minus 3 percent, plus a buyback program over $150 billion. | verified (Nvidia investor page read, 22 November 2026), share derived
- Apple warns about the next far-reaching access right under macOS In the second beta of macOS 27.3 (16 November 2026) Apple urges: anyone granting an app access this far-reaching should understand that it puts every file on the Mac at risk (quoted in translation). Access already granted stays in place; the warning does not reach backwards. | reported (Apple news page read, 22 November 2026)
- Cost per task instead of price per token: Gemini 4, GPT-6.2 and Claude compared On Artificial Analysis (index v5.0.1) one task costs $4.00 with Astra and $26.54 with Claude Fable 5.2 at maximum effort, although both have the same list price per token. Google lists Gemini 4 Hydrogen at $1.50 and $7.50 and Neon at $3 and $15 per million tokens; per Google, Carbon and Argon cost $2 and $10 with a credit until 30 Nov 2026 and $4 and $20 from 1 Dec 2026. Stripe measures a 78.1 percent average score for coding agents, but only 29.4 percent of runs meet all required criteria. | verified (Anthropic and Google price pages, OpenAI system card, Artificial Analysis model pages, Stripe methodology read, 22 Nov 2026)
- AI agent in HR: when the works council must consent Under Regulation (EU) 2026/1744 the AI Act's high-risk duties for hiring and personnel management apply from 2 Dec 2027 (Annex I from 2 Aug 2028). Separately, an agent that processes employee data or can record performance and behaviour often needs the works council's consent in Austria (§ 96, § 96a ArbVG), and in Germany § 87(1) no. 6 BetrVG applies. Not legal advice. | verified (statute texts, EUR-Lex regulation text and Commission page read, 22 Nov 2026)
Consumer hardware
- Librem 5 v2: 100 orders, $1,599, RISC-V and a kill switch Purism reported on 20 November 2026 that the Librem 5 v2 has passed 100 orders; the price is $1,599 and the goal of the campaign is 2,500 devices. The page lists RISC-V as the chipset direction and a hardware kill switch for the radio parts. An early-bird price or tiers are not named there. | reported (Purism post read, 22 November 2026)
- Xiaomi 18 Pro and Pro Max listed in Germany The German product pages are up: the Pro brings 6,000 mAh, the Pro Max 7,050 mAh in silicon-carbon cells, and the pages name a "Snapdragon 8 Elite Extreme Gen 6" as the processor, a 6.9-inch display with a privacy function and a second display on the back. Euro prices are on neither page, so we name none. | reported (product pages read, 22 November 2026)
Software releases
- PostgreSQL 19 is here: rebuild without a lock, plan advice, waiting for a log position Version 19 arrived on 19 November 2026. What matters most for operators:
REPACK CONCURRENTLYreorganizes a table without a write lock,pg_plan_advicegives hints on plan choice,WAIT FOR LSNlets a process wait for a position in the replication log, and asynchronous file I/O is live on several paths. The "nearly doubled" foreign-key check is a project test value; we did not measure it. | vendor figure (release notes read, 22 November 2026) - Rocky Linux 10.3 and AlmaLinux 10.3 with 9.9: a kernel carrying 36 CVE fixes Both rebuilds of Red Hat 10.3 arrived on 16 November 2026. Rocky follows the upstream line; AlmaLinux stresses that its own kernel fixes 36 vulnerabilities still open in the Red Hat 10.3 kernel, and brings Podman 6.1.1 with changes to network and image management. Anyone with Grafana 10 in the estate: legacy alerting is gone in the 12.x releases. | verified (both posts read, 22 November 2026)
- Zulip 13.3 and 12.5 close a critical flaw, with no CVE number The security release of 19 November 2026 fixes what the advisory calls a critical vulnerability (CRITICAL tier, effect on directory authentication) and to this day assigns it no CVE number; the project holds back details. Anyone running the self-hosted variant should apply 13.3 and 12.5 respectively. | verified (GitHub advisory read, 22 November 2026)
- Gitea 28.4.0: 19 vulnerabilities fixed, among them a push race with code execution The release of 15 November 2026 (one day before our reporting window, and the correct state here) carries 19 advisories in its attachment, among them GHSA-6m4f-m9vc-7946 (a race on push up to code execution) and a firmer grip on TOTP brute force. The text has no Security section; this is a patch release with a security effect, and it was never announced as a security update, which the post itself points out. | verified (GitHub release read, 22 November 2026)
- fwupd 2.1.10 brings firmware rollbacks and steadier polling The maintenance release of 16 November 2026 adds rolling back installed firmware, improves the timing with devices that answer slowly, and corrects several small things in the USB path. | verified (GitHub release read, 22 November 2026)
Smart home
- Matter over Thread does not mean Matter over everything: heise tests the Meross MTS155 at 65 euros The test of 17 November 2026 gives the 65-euro thermostat a cheap way in: it hangs on Matter over Thread but maps only basic functions, and for heating schedules and sensor bundles the vendor app stays in play. The article names the Sonoff TRV Gen 2 as the comparison. | reported (heise test read, 22 November 2026)
- How long do updates last? The answer is usually missing WeLiveSecurity gathers the state of play on aging smart-home devices: a 2024 FTC study found that 89 percent of manufacturer pages do not disclose how long products receive updates, and the EU duty on support periods (per product class, mostly at least five years) applies from December 2027. | reported (WeLiveSecurity read, 19 November 2026; visible date missing)
- BADBOX 3.0: a botnet in TV boxes reports 40 billion ad requests a day On 21 November 2026 BleepingComputer and the HUMAN Security blog reported: BADBOX 3.0 injects bid requests through manipulated Android TV boxes, over 4 percent of the ad traffic on television sets and three times the previous record (VASTFLUX); along with it 11 residential proxy networks, the successor build PXYEDGE with roughly 200,000 advertised nodes (peak 340,000), a link to a manufacturer in China, and the takedown of the XCPE infrastructure reported by Google. For anyone running guest wifi or warehouse radio: AOSP devices with no update path are the vehicle here. | verified (both reports read, 22 November 2026)
From the blog
- Two Radeon AI PRO R9700 with Radiance: what Qwen3.8-27B really does on AMD Four engine setups, one box, every figure measured by us; the deep dive is above.
- Mistral's Timestral: the numbers hold, the first place does not GIFT values reproduced to the fourth decimal, and the leaderboard reads differently; deep dive above.
- vLLM 0.34.0 measured: the KV cache default changes, our lane stays as it is Where the new defaults bite and which switch keeps us clear; deep dive above.
- Kiri starts containers fast, the eight-times number only holds on a Mac A short test of the container accelerator: the vendor figure hangs on the Apple platform.
- LiteLLM Customer Management is only in a release candidate, in no stable version Version states read directly instead of a release name taken on faith.
- UniFi Network 11 moves the database, wipes the statistics and has no way back The move to PostgreSQL, the statistics starting empty and the missing way back.
- Optimizely supply-chain attack: where a CSP stops the loader and where it does not Measured on our own three sites, with the check instructions.
- Flipper One on a company network: EUR 11 million in pledges, 339 access ports and no port authentication Crowdfunding numbers and what the device can really do in a switch port.
- One day online, then pulled: NVIDIA's QAD checkpoint of Qwen3.8-Flash-Next Model availability is a state, and a state can be taken back.
- GeForce 621.14 and 621.24 write up to 30 GB an hour to the SSD The constant write path
NVToppsin a self-test with a write tally. - MinIO's AGPL edition is archived, and the free AIStor edition forbids production use Licence text read: "Free" in AIStor does not mean free to use.
- Keycloak 27 is set to drop FGAP V1 and Legacy Token Exchange What the planned break means for integration paths.
- Exchange server retired: the Mailbox Manager runs AD and Microsoft 365 The free path to maintaining mailboxes across Active Directory and Graph together.
- Proxmox 9: the opt-in kernel 7.3 comes from a different source than before Measured at the package source rather than taken from an announcement.
- Microsoft's agent-framework on PyPI: a .pth hook stole credentials Timeline, files, check commands and rotation list, in the digest under Security.
- Cost per task instead of price per token Price lists, Artificial Analysis cost per task and Stripe's evaluation side by side, in the digest under AI and development.
- AI agent in HR: when the works council must consent Austria and Germany side by side, with the state of the high-risk deadlines.
- Inkling 2 Mini on two Radeon cards: speed, context and a PDF extraction Our weekend measurement series, including the numbers the audit did not uphold.
- Token budget per role: LiteLLM keys and teams tested Budget, rate-limit and tag behaviour of LiteLLM 1.110.1 measured without production keys, including the price-0 trap.
- Claude, GPT-6 and Gemini 4 compared 1:1: 12 models, prices and measurements List prices, Artificial Analysis cost per task and what the LiteLLM price file says about the December price jump for Gemini 4.
Sources for this edition
- Deep dive R9700/Radiance: Post on two Radeon AI PRO R9700 with Radiance, Level1Techs video on the R9700 (video, not evaluated)
- Deep dive Timestral: Post on recounting Timestral, GIFT-Eval board, Mistral announcement, repository
- Deep dive vLLM: vLLM 0.34.0 release, post on vLLM 0.34.0 measured
- More sources, digest: GTIG AI Threat Tracker, FBI FLASH on ARTEX, heise on ARTEX and the EU Commission, Oracle CSPU November 2026, Calif on Java4Shell, NGINX advisories, CISA KEV, CISA KEV as a JSON feed, LiteLLM GHSA-4xm2-x5g5-wc8q, Project Zero policy, Golem on NSAnet, heise on US banks and the NSA, heise on the Exchange post-incident report, Microsoft Security Blog on the phishing wave, WDR on Phantasialand, Windows Server 2028 on Learn, MissionControlAggregator analysis, Windows Insider build announcement, microsoft.ai on MAI weights, Azure Cobalt 200, Born's IT on the boot loop, Born's IT on the activation error, Born's IT on KB5131085, MSRC CVE-2026-124223
- Sources of the added posts: GitHub advisory GHSA-764m-3ppc-5m73, Wiz on agent-framework, Aikido on agent-framework, BleepingComputer on agent-framework, Post: cost per task, Post: works council and AI agent, Post: Inkling 2 Mini on two Radeon cards
- Hardware sources: Nvidia Rubin HPC, Nvidia Vera for science, Lambda on CPO, Nvidia post on Lambda photonics, Intel D270, StorageReview on Diamond Rapids
- AI sources: OpenAI GPT-6.2 safety page, Simon Willison's take, Qwen4-Max blog, Hugging Face on OpenRouter, Nvidia on Hugging Face, Nvidia Q3 FY27, Apple news on macOS 27.3 beta
- Consumer, release and smart home sources: Purism Librem 5 v2, Xiaomi 18 Pro, Xiaomi 18 Pro Max, PostgreSQL 19 release notes, Rocky Linux 10.3, AlmaLinux 10.3 and 9.9, Zulip advisory, Gitea 28.4.0, fwupd 2.1.10, heise on the Meross MTS155, WeLiveSecurity on update durations, BleepingComputer on BADBOX 3.0, HUMAN Security blog on BADBOX 3.0
- Logos (CC0): Simple Icons 16.36.0, file amd.svg, Simple Icons 16.36.0, file mistralai.svg, Simple Icons 16.36.0, file vllm.svg. Brand guidelines: amd.com, mistralai.com, vllm-project.org.
Compiled Monday, 23 November 2026. Sources: Google Threat Intelligence Group, Mandiant, FBI IC3, heise online, Oracle, Calif.io, NGINX, CISA, NVD, GitHub (BerriAI, zulip, go-gitea, fwupd, vllm-project), Project Zero, Microsoft (Security Blog, MSRC, Learn, Windows Insider, Azure), Borncity, Golem, WDR, Nvidia, Lambda, Intel, StorageReview, OpenAI, Simon Willison, Alibaba/Qwen, Hugging Face, Apple, Purism, Xiaomi, PostgreSQL, Rocky Linux, AlmaLinux, WeLiveSecurity, BleepingComputer, HUMAN Security, Simple Icons. Whatever comes after it goes into the next edition.
senn-tech