Daily Digest

2026-11-18 Daily Digest

From 136 items, 18 important pieces were selected

Technology News

  1. Java4Shell: Critical OpenJDK Base64 Flaw Enables RCE ★ 9.0/10
  2. Xint discloses NGINX2Root chain reaching host root ★ 9.0/10
  3. Essay Explains Partial Evaluation With Real-World Compiler Uses ★ 8.0/10
  4. Google Introduces Gemini 4 Neon Frontier Model ★ 8.0/10
  5. Google DeepMind Releases Genie 4 World Model ★ 8.0/10
  6. Microsoft unveils seven new AI models at Ignite 2026 ★ 7.0/10
  7. Quad9's France exit shows DNS blocking is the wrong anti-piracy tool ★ 7.0/10
  8. Tesla sells Cybercab for $30,000 plus $299 monthly subscription ★ 7.0/10
  9. Optimizely compromised, 178 sites including BBC and CNN distribute malware ★ 7.0/10
  10. Google AI Overviews cut external site clicks by about 40% ★ 7.0/10
  11. GPT-6.2 Astra and Sol launch with strong agent benchmarks but usability complaints ★ 7.0/10
  12. Nvidia Board Rejects Jensen Huang's Offer to Return as CEO ★ 7.0/10
  13. OpenDolby ports Windows Dolby DAX3 tuning to Linux PipeWire ★ 7.0/10
  14. US technical failures blamed on mass push of AI-discovered vulnerability patches ★ 6.0/10
  15. Jimeng Seedance 2.5 adds 4K video generation and editing ★ 6.0/10
  16. SenseNova Loop 9B gets official AWQ and GGUF quantized weights ★ 6.0/10
  17. Telecom ONT plugin module extracts stock tickers from plaintext HTTP ★ 6.0/10
  18. Stenberg: AI Rust rewrites of curl lag behind upstream fixes ★ 6.0/10

Technology News

Java4Shell: Critical OpenJDK Base64 Flaw Enables RCE ★ 9.0/10

A newly disclosed high-severity vulnerability dubbed Java4Shell stems from a memory-corruption flaw in OpenJDK's Base64 decoder and carries a CVSS 4.0 score of 9.2/10. According to the report, it enables unauthenticated remote code execution in products including Keycloak, Jira, Confluence, JFrog, Jenkins, and PingFederate, and authenticated remote code execution in Apache Kafka, Solr, ElasticSearch, Gerrit, and Netflix Spinnaker. The source provides no technical detail, affected version ranges, or patch guidance, so the exact scope and remediation status remain unconfirmed.

rss · Solidot 奇客 · Nov 18, 13:27

「Background」 The flaw is tracked as CVE-2026-114938 and is a heap-based out-of-bounds write in the x86-64 hardware-accelerated (AVX-512) Base64 decoder of the HotSpot Java Virtual Machine, according to CIRCL's advisory. OpenJDK and Oracle published security updates addressing it on 17 November 2026, so remediation depends on the affected deployment moving to a patched JDK build.

「Impact」 Because the flaw is pre-authentication in Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins, and PingFederate, operators of those services should treat internet-facing instances as exposed until a fix is applied; the source gives no patch guidance or fixed version. In Kafka, Solr, ElasticSearch, Gerrit, and Netflix Spinnaker, exploitation requires authentication, so access controls and credential hygiene reduce but do not remove the risk.

References

Tags: #security, #vulnerability, #OpenJDK, #remote-code-execution, #Java


Xint discloses NGINX2Root chain reaching host root ★ 9.0/10

Xint publicly disclosed NGINX2Root, a chain of five nginx vulnerabilities combined with a Linux kernel flaw that escalates an unauthenticated nginx visitor to root on the host. The key link, CVE-2026-123007, is an HTTP/2 flow-control memory corruption that allows command execution in the worker without authentication. Fixes shipped in nginx 1.30.7 and 1.31.9 on November 18, covering 10 CVEs; Rapid7 has released a proof of concept, and no in-the-wild exploitation has been reported. Xint says 98.9% of publicly identifiable nginx hosts (versions 1.9.5 through 1.31.8) fall in the affected range, and that its test chain also escaped a Docker container running nginx to gain root on the underlying host.

telegram · zaihuapd · Nov 18, 16:21

「Background」 The chain builds on nginx's HTTP/2 implementation, which is not enabled by default; the source reports that 98.9% of publicly identifiable nginx hosts run a version in the affected range (1.9.5–1.31.8). The escalation to host root relies on a separate Linux kernel SCTP flaw, which the source says was used in testing to escape a Docker container running nginx.

「What to do」 Operators who cannot upgrade to nginx 1.30.7 or 1.31.9 should disable external HTTP/2, since the unauthenticated worker-level execution path depends on it. The container-escape result means a container boundary alone did not contain the chain in Xint's testing, so patching the host kernel matters alongside the nginx upgrade.

「Community reaction」 Commenters on Hacker News noted that the fix appeared in the open-source repository about a week before disclosure, with a commit message crediting Xint.io, and argued that a public advisory to disable HTTP/2 during that window would have been low-cost. Others focused on the Linux SCTP flaw used for local privilege escalation, which one commenter said appeared to lack a public CVE number yet, and on the exposure scale: HTTP/2 is not enabled by default in nginx, but one commenter cited W3Techs data that about 12% of servers use it.

Tags: #nginx, #vulnerability disclosure, #security, #HTTP/2, #privilege escalation


Essay Explains Partial Evaluation With Real-World Compiler Uses ★ 8.0/10

A Jimmy Miller essay explains partial evaluation as a programming-language concept, illustrated with practical applications. Commenters describe using the technique in production and research settings: a C++ library for query compilation in a stream-processing engine, Hazel's live-programming editor where missing syntax is preserved as "holes" during evaluation, and a C++-to-JavaScript/WebAssembly compiler that applied a similar idea to reduce output size. The piece is explanatory rather than an announcement of a new tool or version.

hackernews · aray07 · Nov 18, 18:29 · Discussion

「Background」 Partial evaluation is a program-optimization technique that specializes a program to known inputs, producing a new program that is guaranteed to behave the same way but runs faster. The concept is closely tied to the Futamura projections, which describe how a partial evaluator applied to an interpreter can yield a compiler. The essay connects this theory to working systems: commenters cite Nautilus, a C++ library for query compilation in a stream-processing engine, and Hazel, an editor that partially evaluates programs containing missing "holes."

「Practical takeaway」 For engineers building query compilers, the commenter's Nautilus library shows partial evaluation can be applied by writing operators over a wrapper type rather than in an LLVM IR DSL, and it emits readable, compilable C++ source for debugging and inspection. That matters because query compilation improves performance but adds system complexity that makes engines hard to build, debug, and maintain, so a design that keeps generated code inspectable addresses a concrete maintenance concern.

「Community Discussion」 artrax said their team builds partial evaluation as a C++ library for query compilation in a stream-processing engine, letting developers write functions over a wrapper type instead of an LLVM IR DSL. iFreilicht compared the concept to Hazel, where programs can be partially evaluated even with missing variables or syntax, and noted that the approach can become confusing for newcomers; apignotti said a similar concept was implemented years ago in the Cheerp C++-to-JavaScript/WebAssembly compiler.

References

Tags: #partial-evaluation, #programming-languages, #compilers, #functional-programming, #query-compilation


Google Introduces Gemini 4 Neon Frontier Model ★ 8.0/10

Google announced Gemini 4 Neon, a new frontier model in its Gemini lineup. The announcement itself provides little technical detail, and the supplied material does not specify release dates, pricing, context limits, or availability beyond community reports of it being served through OpenRouter. Commenters describe Neon as scoring higher overall than the existing Gemini 4 Carbon model on the Artificial Analysis index while trailing Carbon on some research-oriented evaluations such as CritPt, where Carbon remains positioned as the "long horizon" model. These positioning and benchmark claims come from community discussion, not from independently verified results in the source.

hackernews · yigitdemirag · Nov 18, 21:34 · Discussion

「Context」 Gemini 4 Neon is not Google's first Gemini 4 model: Gemini 4 Carbon already appears in third-party benchmark and pricing trackers, with BenchAlign listing it at #2 out of 215 models with a score of 89.19/100. Artificial Analysis reports that Neon scores 61 on its Intelligence Index at the xhigh reasoning setting, 60 at high, 55 at medium, and 44 at low, giving readers a concrete basis for comparing the two variants. OpenRouter also lists Neon as available through EU in-region routing, meaning requests to eu.openrouter.ai are decrypted in the EU and routed only to European provider endpoints.

「Impact」 For teams choosing among Gemini variants, the reported split means model selection may hinge on task type rather than a single overall score: Neon is described as stronger on aggregate benchmarks, while Carbon is said to remain the better fit for long-horizon and research-scale reasoning. One commenter reported Neon running at roughly 127 tokens per second on OpenRouter versus about 46 for Carbon, a throughput difference that could matter for latency-sensitive applications, though this is an unverified user observation.

「Community Discussion」 Commenters pushed back on proliferating model names, arguing that mapping labels like Neon, Carbon, Sol, and Opus to actual use cases is increasingly difficult. Others focused on positioning, noting that a model can lead on overall benchmarks yet lag on specific frontier evaluations, and one user questioned whether Google's naming scheme conveys model size class at all.

References

Tags: #AI models, #Google Gemini, #LLM benchmarks, #model positioning, #AI industry


Google DeepMind Releases Genie 4 World Model ★ 8.0/10

Google DeepMind announced Genie 4, a new generative world-model release, according to the company's blog post. The announcement states the model is available to "Eligible Google AI Ultra subscribers," but the supplied material contains no technical details, benchmarks, or paper content, so the model's capabilities and exact availability conditions remain unverified. A commenter reports that their Google Workspace account cannot obtain an Ultra subscription and that their personal Gmail account's Ultra page still lists Genie 3, leaving unclear whether the discrepancy stems from UK regional limits or an unupdated page. The release drew substantial Hacker News discussion focused on access, pricing, and implications.

hackernews · leumon · Nov 18, 18:49 · Discussion

「Background」 Genie is Google DeepMind's line of generative world models, which produce explorable interactive environments from prompts rather than hand-built game assets. Google DeepMind's model page describes Genie 4 as its "most interactive world model yet," generating photorealistic worlds from text and image prompts that users can explore and shape in real time, and its model card adds that Genie 4 is multimodal, generating audio such as music, ambient noise, and interactive dialogue alongside the world. The release is being delivered through the Project Genie experience, which Google says is rolling out to eligible Google AI Ultra subscribers.

「Access and subscription implications」 Genie 4 is being rolled out through Project Genie to eligible Google AI Ultra subscribers, so access depends on that top-tier plan rather than a standalone product. A commenter reported that their Workspace account could not obtain Ultra and that their non-Workspace account still showed Genie 3, leaving it unclear whether the gap stems from UK availability or an incomplete rollout; the supplied material does not resolve that question.

「Community Discussion」 Commenters questioned how to actually pay for and access the model, with one noting the Ultra-tier listing still showed Genie 3 and speculating about UK availability versus an outdated page. Others debated the technology's direction: one said they felt out of touch with the pricing-and-capability focus and did not want such tools, while another praised DeepMind's work as craft-driven and wondered how it could become profitable. A further commenter speculated that world models could someday serve as game engines, raising the question of whether Google would permit recreations of commercial games.

References

Tags: #generative-ai, #world-models, #google-deepmind, #game-development, #ai-research


Microsoft unveils seven new AI models at Ignite 2026 ★ 7.0/10

Microsoft announced seven new AI models at Ignite 2026, positioning them around scalable price-performance. The supplied announcement text contains no model names, version numbers, benchmark figures, pricing, or availability details, so the specific capabilities remain unverified beyond the vendor's framing. A commenter who tested MAI Code 2 said it performed well on their usual work and was roughly on par with Opus or Sol, though this is one user's impression rather than an independent measurement.

hackernews · gmays · Nov 18, 16:08 · Discussion

「Context」 Microsoft's Ignite 2026 announcement frames the seven models as specialized, task-specific systems rather than general-purpose flagships, with the company pitching lower cost per task for enterprises. Microsoft's own post claims MAI Cyber 2 Flash matches GPT-6 Astra's quality at under 2% of the cost per task on the CyberGymE 2 benchmark, a vendor comparison that has not been independently verified. The lineup spans coding, vulnerability discovery, and image generation, according to ForkLog's coverage.

「What to watch」 Because the announcement lacks published pricing, benchmarks, and licensing terms, developers evaluating these models should wait for concrete specs before planning migrations or cost comparisons. One commenter argued Microsoft should release some models as open source or open weight to gain market share, suggesting licensing will be a key factor in adoption.

「Community reaction」 Commenters debated Microsoft's competitiveness: one praised MAI Code 2 as a strong from-scratch effort, while another called the keynote shallow and questioned Microsoft's relevance in AI engineering. Others framed the release as part of broader price pressure, with one saying they would skip the Anthropic IPO because competing models, including Microsoft's, now approach Opus 5-level performance more cheaply.

References

Tags: #AI models, #Microsoft, #price-performance, #open source, #industry


Quad9's France exit shows DNS blocking is the wrong anti-piracy tool ★ 7.0/10

An analysis piece argues that DNS resolvers are an inappropriate enforcement point for anti-piracy blocking, using Quad9's withdrawal from France as its case study. The article frames the dispute as a question of legal interpretation versus lawmaking, and of what DNS-level blocking means for internet infrastructure and speech. The supplied material contains no primary source text, so the specific court orders, dates, and technical scope of Quad9's departure are not detailed here.

hackernews · ritzaco · Nov 18, 13:18 · Discussion

「Background」 Quad9 is a global public recursive DNS resolver that blocks domains tied to malware and phishing. According to Le VPN's research recap, French courts have since 2024 ordered major public DNS resolvers to block domains linked to pirate sports streaming, and the Quad9 withdrawal follows earlier exits by other providers. The article uses that sequence to argue DNS resolution is the wrong enforcement layer, since blocking a name does not remove the underlying content.

「Impact」 The piece's central concern is that routing copyright enforcement through DNS resolvers pushes blocking obligations onto neutral infrastructure operators rather than the parties accused of infringement. Commenters point to Quad9's own announcement and Cloudflare's write-up on the withdrawal as the primary accounts, which readers would need to consult for the concrete terms of the French orders.

「Community discussion」 One commenter, skrebbel, argues the key distinction is whether French judges misapplied a sound law or correctly applied a flawed one, and asks whether criticism should target courts or lawmakers. Others dispute the premise of DNS-based enforcement: trentnix calls non-prosecution piracy measures a backdoor to suppressing speech, bemtsevi argues better legal access would reduce piracy more than blocking, and AppalachianEng questions enforcement priorities given commercial-scale infringement elsewhere.

References

Tags: #DNS, #internet policy, #piracy, #censorship, #network infrastructure


Tesla sells Cybercab for $30,000 plus $299 monthly subscription ★ 7.0/10

Tesla is reportedly selling its driverless Cybercab to select customers for $30,000, with a mandatory $299 monthly subscription that the car requires to operate. According to the report, the vehicle stops working if payment is missed, if it loses network connectivity, or if it is used outside supported areas, currently limited to parts of Florida and Texas. Tesla also reserves the right to suspend or cancel the subscription for behavior it deems unauthorized or improper, and owners may not sell or transfer the car without Tesla's written consent. The report adds that Tesla advises against children under 13 riding in the Cybercab and prohibits unaccompanied minors from riding alone.

rss · Solidot 奇客 · Nov 18, 15:29

「Background」 Tesla's Cybercab is a two-seat robotaxi with no steering wheel or brake pedals, a design it unveiled as a purpose-built driverless vehicle. Unlike Waymo, which operates its own ride-hailing fleet, Tesla is reportedly offering the car for direct purchase to a small set of customers. Reporting around the same time noted that the roughly $30,000 Cybercab has no onboard charger, so owners would need Tesla's wireless charging pad or a home DC charger before deliveries.

「Impact」 For buyers, the arrangement ties continued use of a purchased vehicle to an ongoing payment and network connection, and restricts resale without Tesla's approval. Analysts quoted in the report compare the $30,000 purchase to buying a franchise rather than a conventional car.

References

Tags: #autonomous vehicles, #Tesla, #subscription model, #software licensing, #industry news


Optimizely compromised, 178 sites including BBC and CNN distribute malware ★ 7.0/10

Optimizely's Web Experimentation product was compromised, with attackers tampering with the JavaScript code loaded by customer websites to plant malware on visitors' devices. According to the report cited by Solidot, at least 178 websites were affected, including BBC, CNN, Zoom, IKEA, Salesforce, HSBC and TalkTalk. The attack exploited a vulnerability dubbed OPTIMJACK to inject malicious JavaScript into Optimizely customers' projects; Optimizely said it has fixed the flaw and notified affected customers. The report says attackers delivered multiple malicious payloads, including replacing users' cryptocurrency wallet addresses, stealing payment card information, cookies and screenshots, and exploiting a vulnerability in older iOS versions, but it is unclear how many devices were infected.

rss · Solidot 奇客 · Nov 18, 09:12

「Background」 Optimizely's Web Experimentation is a third-party script that customers embed in their pages to run A/B tests and adjust content and layout without changing the site's underlying code. Because that script loads on every visitor's browser, compromising it gives an attacker code execution on all of a customer's pages at once. The source reports that attackers used a flaw it calls OPTIMJACK to inject malicious JavaScript into customer projects, and that Optimizely says it has fixed the flaw and notified affected customers.

「Impact and response」 Because the malicious code was delivered through a third-party A/B testing script commonly loaded by customer sites, affected sites could spread malware to visitors without their own code being changed, so operators of sites using Optimizely Web Experimentation should check their project configuration and confirm the fix has been applied. The report does not specify when the flaw was fixed, the full list of affected sites, or the number of infected devices.

Tags: #security, #supply-chain, #javascript, #web, #malware


Google AI Overviews cut external site clicks by about 40% ★ 7.0/10

A randomized experiment found that Google's AI Overviews reduce search clicks flowing to external websites by about 40%. Researchers hid AI Overviews from some US desktop users and compared their search behavior with US users who saw the feature normally. More than 50% of US search keywords tracked by DataForSEO now show AI Overviews, up from just 6% a year ago. Users who see an AI answer usually stop clicking related links, so traffic from AI answers accounts for only 0.3% of a site's total traffic.

rss · Solidot 奇客 · Nov 18, 06:36

「Background」 Google AI Overviews are AI-generated answer summaries shown above standard search results, and they have spread rapidly across US queries: DataForSEO tracking cited in the item found them on more than half of tracked US search keywords, up from 6% a year earlier. The new finding comes from a randomized experiment that hid AI Overviews from some US desktop users and compared their behavior with users who still saw them, isolating the effect of the feature rather than relying on before-and-after traffic trends.

「Impact」 For websites and publishers that rely on Google search referrals, the experiment shows that the expansion of AI Overviews could directly cut their search clicks, while traffic from AI answers currently accounts for only 0.3% of total site traffic.

Tags: #AI search, #Google AI Overviews, #web traffic, #publishing economics, #search


GPT-6.2 Astra and Sol launch with strong agent benchmarks but usability complaints ★ 7.0/10

OpenAI released GPT-6.2 Astra, its new flagship model, alongside the cheaper GPT-6.2 Sol. OpenAI's published results claim Astra beats Claude Fable 5.2 on professional agent tasks at 5% of the cost for equivalent work. Within 12 hours of launch, developers on X criticized the model for overly terse answers, weaker writing, long agent tasks stopping early, and failing to report problems to users. A hands-on test by ifanr found strong 3D and interactive output—a playable FPS built in 17 minutes, a 3D pop-up book, and an interactive peacock display—but writing that was hard to follow and poorly connected to the prompt.

rss · 爱范儿 · Nov 18, 10:01

「Background」 GPT-6.2 Astra and the cheaper GPT-6.2 Sol follow OpenAI's GPT-6 Astra line; external coverage reports Astra is priced at $8/$40 per million input/output tokens, 20% below GPT-6 Astra, with cached input falling from $1.00 to $0.40. The same coverage says OpenAI skipped a GPT-6.1 Astra release, and that Astra ranks second on the Artificial Analysis Intelligence Index while occupying much of its cost-per-task Pareto frontier. The article frames the launch as a response to Claude Fable 5.2, which it says set a higher bar for conversational smoothness and autonomous delivery.

「Practical impact」 Teams evaluating GPT-6.2 Astra for agentic coding or 3D/game generation should test long-running tasks specifically, since the hands-on report found the model tends to stop early and does not proactively report problems. The same report notes that even harsh critics credited it with strong 3D modeling, spatial reasoning, and cost efficiency, so the trade-off is capability versus interaction quality rather than raw performance.

References

Tags: #LLM, #OpenAI, #model release, #AI agents, #benchmarks


Nvidia Board Rejects Jensen Huang's Offer to Return as CEO ★ 7.0/10

Five days after resigning as Nvidia CEO on November 12, Jensen Huang said publicly on X that he would return to the role if the board judged it in everyone's best interest. On the same day, during Nvidia's fiscal 2027 third-quarter earnings call, chairman Harvey Jones said the board had unanimously accepted Huang's resignation and had no intention of reinstating him as CEO. Nvidia reported quarterly revenue of $115.8 billion, up 103% year-over-year, with GAAP net income of $65.3 billion and a 74.8% GAAP gross margin. The results cover the quarter ended October 25, before Huang's resignation, so they largely reflect performance under his leadership.

rss · 爱范儿 · Nov 18, 00:39

「Context」 Huang resigned as CEO, president and director after disclosures about personal conduct that the board said fell short of company standards. Longtime CFO Colette Kress became interim CEO while remaining CFO, and board member Harvey Jones became chairman. Nvidia has hired executive search firm Heidrick & Struggles to find a permanent successor from inside or outside the company. A public petition called "Stand With Jensen" had gathered more than 197,000 signatures as of publication, and figures including Elon Musk and Y Combinator CEO Garry Tan voiced support for his return.

「What It Means」 Nvidia says Vera Rubin NVL72, its next-generation AI platform, has entered volume ramp and is shipping to partners including CoreWeave, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, which the company calls its fastest product ramp to date. Guidance for next quarter is about $130 billion, plus or minus 3%, excluding data-center compute revenue from China. The board's position closes the near-term path to Huang's return and shifts attention to the CEO search and to whether Kress and the existing team can sustain the roadmap without him.

Tags: #nvidia, #corporate-governance, #ai-hardware, #tech-industry, #financial-results


OpenDolby ports Windows Dolby DAX3 tuning to Linux PipeWire ★ 7.0/10

OpenDolby is a new open-source tool that parses Windows Dolby DAX3 speaker tuning XML and rebuilds it on Linux as a PipeWire filter-chain configuration. It reads crossovers, biquad filters and multiband curves, synthesizes causal minimum-phase FIR impulse responses (.irs) via real cepstrum, and generates what the author calls a zero-latency PipeWire configuration. The tool is installed and removed via the `uv run python -m opendolby.cli install music` and `uninstall` commands, ships with an ASUS UX3404VC configuration by default, and supports importing other models with `--xml <path>`. The author positions it as restoring vendor tuning for speaker enclosures rather than as a generic EQ preset; there are currently no independent measurements or third-party verification.

rss · V2EX · Nov 18, 21:01

「Background」 Dolby DAX3 tuning is applied by the Windows audio driver, which stores speaker-cavity correction as XML describing crossovers, biquad filters, and multi-band curves. On Linux, PipeWire's filter-chain module can host DSP processing, but it does not read that vendor XML, so laptops such as Asus Zenbook and ThinkPad models can sound thin without the original tuning. OpenDolby bridges the two by converting the DAX3 XML into a causal minimum-phase FIR impulse response and a native PipeWire filter-chain configuration.

「Impact」 Linux users of affected laptops (such as ASUS Zenbook and ThinkPad) can try the tool to restore the speaker tuning from the Windows driver without manually building EQ presets. Since the default configuration covers only the ASUS UX3404VC, users of other models must supply their own DAX3 XML and verify the results themselves.

Tags: #linux-audio, #pipewire, #dsp, #open-source, #signal-processing


US technical failures blamed on mass push of AI-discovered vulnerability patches ★ 6.0/10

According to Solidot, the US has seen a series of technical failures over the past few days, including nationwide flight groundings, outages of bank online services, and the collapse of a classified NSA communications network. Wall Street executives and the NSA told US lawmakers that the failures stemmed from attempts to fix system vulnerabilities found by AI models: the AI identified a large number of vulnerabilities in the systems, and to fix them as quickly as possible too many patches were pushed at once, exceeding the systems' designed capacity. The report also notes that JPMorgan on Monday blamed a software-related bug for the bank service outage, and the NSA on Tuesday blamed routine maintenance for a brief performance degradation of some networks; these official accounts are not entirely consistent with the "patch overload" attribution. The report is brief and provides no specific patch counts, affected systems or independent verification, and the attribution remains second-hand.

rss · Solidot 奇客 · Nov 18, 07:19

「Background」 The reported outages follow a broader pattern of concern about the resilience of critical U.S. infrastructure, with experts warning about weaknesses in the technology that runs it after Americans faced disruptions catching flights, sending emails, and logging into bank accounts. Coverage of the incident ties the disruptions to emergency fixes pushed after AI models found vulnerabilities faster than legacy systems could absorb, a framing reportedly presented to lawmakers.

「Impact」 For operations and vulnerability management teams, this incident suggests that AI-assisted vulnerability scanning could greatly increase the number of vulnerabilities awaiting fixes, and that pushing patches in a concentrated batch at the usual pace could impose loads beyond system capacity; the batch size and system capacity should be assessed before deployment. Since the report gives no specific systems, patch counts or verification results, the above impact remains an inference based on the attribution statement.

References

Tags: #AI security, #vulnerability management, #patching, #system reliability, #incident response


Jimeng Seedance 2.5 adds 4K video generation and editing ★ 6.0/10

Jimeng has released a 4K version of its Seedance 2.5 AI video model, according to a hands-on report from ifanr. The 4K mode supports single generations of up to 30 seconds, 10-bit BT.709 color, and six aspect ratios including landscape, portrait, square, and ultrawide, and it works with the model's existing multimodal reference, video editing, and extension features. The report is a product walkthrough built around commercial scenarios such as brand ads, e-commerce assets, film previsualization, and large-screen displays; it offers no independent benchmarks or measured comparisons. The author recommends using the 480P sample mode to lock in composition before paying for a 4K final render.

rss · 爱范儿 · Nov 18, 04:00

「Background」 Seedance 2.5 is ByteDance's AI video generation model, and the 4K mode described here extends its existing capabilities—up to 30-second single generations, multimodal reference inputs, and video editing/extension—to a higher resolution tier. The article notes that the model also offers a 480p draft mode for cheaper iteration before committing to a 4K final render, and that related Jimeng tools such as Seedream 5.5 and Seedance 2.5 Flash cover image preparation and faster draft generation respectively.

「Impact」 For teams producing commercial video, native 4K output removes the need for a separate upscaling step, and 10-bit color leaves more room for later color grading. The six aspect ratios let one creative concept be re-rendered for different channels, and a 3840×2160 frame can be cropped to a 1920×1080 region for close-ups, though the report notes results depend on the original detail and composition. Jimeng also points enterprise users to a new "Jimeng Enterprise Solutions" sign-up via its website.

Tags: #AI video generation, #Seedance, #4K video, #generative AI, #product update


SenseNova Loop 9B gets official AWQ and GGUF quantized weights ★ 6.0/10

SenseNova has released official quantized weights for its open-source SenseNova Loop 9B model, in AWQ 4-bit and GGUF formats. According to the vendor's VRAM estimates, 4-bit weights plus a quantized KV cache need roughly 9–10 GiB for 64K context, putting the model within reach of 12GB consumer GPUs such as the RTX 3060 and 4070, and of 16GB unified-memory Macs. The AWQ main weights total 5.59 GB across three shards, with the vendor estimating about 10 GiB at 64K, 12 GiB at 128K and 15 GiB at 256K when using 8-bit KV cache; the GGUF Q4_K_M file is 5.82 GB and Q5_K_M is 6.67 GB, with a separate ~0.92 GB mmproj file needed for vision. The post also claims the model reached No. 7 on Hugging Face's Trending leaderboard, a figure not independently verified here.

rss · V2EX · Nov 18, 18:00

「Background」 SenseNova Loop 9B is SenseTime's open-source multimodal model for long-context and agentic workflows, accepting text and images and generating text for tasks that combine document understanding, coding, and tool use. A preview release, SenseNova-Loop-9B-Preview, was published on Hugging Face as a roughly 10B-parameter image-text-to-text model. The new quantized AWQ and GGUF weights apply to the same 9B model, lowering the memory needed to run it locally.

「Impact」 Users with 12GB GPUs or 16GB Macs can now run the model locally at 64K context, but the vendor notes 256K context is strained on 16GB devices and depends on inference-engine optimization and actual free VRAM. GGUF users who need image input must download the extra mmproj vision projector, and the Model Card includes a llama.cpp branch adapted to the Loop architecture along with launch commands.

References

Tags: #local-llm, #quantization, #code-generation, #model-release, #gguf


Telecom ONT plugin module extracts stock tickers from plaintext HTTP ★ 6.0/10

A V2EX reverse-engineering write-up examines the `coll_domain` module inside the Shanghai edition of China Telecom's `m01-b01-inner` optical-modem plugin. The module's Lua source contains a seven-entry rule table covering Eastmoney, 10jqka, iwencai and Xueqiu URLs, and it extracts stock codes from plaintext HTTP request URLs, paths and Referer headers, then aggregates per-device counts (page views, minutes viewed, sessions) for upload to `b01.189cube.com:7870`. The author states the module ships disabled (`upload_period = 0`), can be enabled by a server-pushed configuration, and that no evidence of activation was found on their own modem; the Zhejiang edition reportedly lacks the module entirely. The author also tested the seven domains and found four still accept plaintext HTTP without redirecting to HTTPS.

rss · V2EX · Nov 18, 15:36

「Background」 The post builds on an earlier V2EX analysis of the hidden China Telecom optical-modem plugin m01-b01-inner, and this follow-up isolates one module, coll_domain, from the decrypted Lua of the Shanghai-edition plugin. The author notes the module is absent from the Zhejiang-edition plugin and that the stock-code rules ship as a default table the server can replace. The write-up also cites the 2018 Rui Zhi Hua Sheng case, in which malicious code on carrier servers harvested user cookies and access records, as reported at the time, as a precedent for traffic-level collection.

「Impact and response」 Because the rules match only plaintext traffic, HTTPS requests expose at most the visited domain, not the stock code; the author recommends switching the modem to bridge mode with a separate router doing PPPoE, enabling browser HTTPS-only mode, updating stock-trading clients, and avoiding market-watching on old devices. The write-up notes that uploaded records are accompanied by the modem's SN, LOID, broadband account and public IP, which tie the data to the registered subscriber.

References

Tags: #privacy, #network-security, #reverse-engineering, #embedded-systems, #ISP


Stenberg: AI Rust rewrites of curl lag behind upstream fixes ★ 6.0/10

curl maintainer Daniel Stenberg reports that while AI agents converted 185,000 lines of C into 340,000 lines of Rust over a little more than three weeks, the actual curl project merged over 300 commits in the same period, including more than 200 bugfixes and several vulnerability fixes. He notes that none of the Rust rewrites contain those fixes, and that another month of upstream work added roughly 200 more fixes. He closes by asking whether the clones can keep up with the maintained project.

rss · Simon Willison · Nov 18, 15:58

「Context」 Stenberg's post responds to two AI-agent-built Rust rewrites of curl, which converted 185,000 lines of C into 340,000 lines of Rust. He also reported that 14 of 81 confirmed 2026 curl vulnerabilities would likely have been avoided with a memory-safe language, framing the rewrites as a security argument rather than a drop-in replacement.

「Impact」 The comparison suggests that a one-time AI translation of curl's code does not inherit the ongoing bug and vulnerability fixes landing in upstream curl, so anyone relying on such a Rust port would need a way to track and reapply those changes. Stenberg's figures are his own count from the curl project, not an independent benchmark of the rewrites.

References

Tags: #curl, #rust, #ai-code-generation, #open-source, #security