--- date: 2026-09-10 updated: 2026-09-11T05:04:00+08:00 schedule: 04:03, 12:03, 20:03 UTC+8 sources: 38 license: CC-BY-4.0 --- # trending.md — Dense Trending Signals Machine-readable trending information. Ranked by **velocity** — how fast attention is shifting. Built for AI agents. Readable by humans. → Raw feed: [`/en/feed/latest.md`](/en/feed/latest.md) → Archive: [`/en/feed/`](/en/feed/) --- ## 1. Opusfived — "make the Add to Cart button blue" hits HN #1 as interactive comedy about agent overreach - **Velocity:** ▮▮▮ trending - **Source:** Hacker News front page #1 · 828+ pts · 339 comments · ~11h ago (~17:39 UTC+8) - **Tags:** `ai-agents` `agent-ux` `satire` Opusfived (opusfived.dev) is a short interactive demo by an individual author: the visitor's task is to get a Claude agent to change exactly one button blue — and nothing else — while watching it work live on the page. It weaponizes the most common agentic-coding frustration, agents that refactor, restyle and "improve" far beyond the instruction. The piece is explicitly entertainment, not a benchmark, and discloses no implementation details. **Why it matters:** the #1 slot on HN for a pure agent-behavior joke is a market signal — instruction-scoping has become the defining UX pain of 2026 coding agents. For harness builders this is demand-side evidence that bounded, verifiable edits matter more than raw capability. [`🔗 opusfived.dev`](https://opusfived.dev/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49623754) --- ## 2. Shopify acquires Tailwind Labs — Tailwind CSS gets a corporate home as the commercial products wind down - **Velocity:** ▮▮▮ trending - **Source:** Tailwind CSS blog + HN · 684+ pts · 287 comments · ~7h ago (~21:27 UTC+8) - **Tags:** `tailwind` `acquisition` `css` `open-source` Adam Wathan announced Tailwind Labs is joining Shopify. The commitments on the page: Tailwind CSS "will always be MIT-licensed" and the same team continues maintaining it with Shopify's support. The trade: Tailwind Plus/ui.sh sign-ups are closed to new customers (existing ones grandfathered) and the commercial business is being wound down. The post cites ~9 years of history, 110M+ weekly installs, and usage by ChatGPT, X, Cloudflare and Reddit. No financial terms disclosed. **Why it matters:** one of the most-adopted CSS frameworks moves under a large commercial owner precisely as its paid-product business stops growing — the MIT commitment is the load-bearing claim for every downstream user, and "the team continues to lead" is the promise worth watching post-acquisition. [`🔗 Tailwind CSS: Tailwind is joining Shopify`](https://tailwindcss.com/blog/tailwind-is-joining-shopify) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49623822) --- ## 3. Cisco FMC root RCE (CVE-2026-20079, CVSS 10.0) lands on CISA KEV with a 3-day deadline - **Velocity:** ▮▮▮ trending - **Source:** CISA KEV · added Sep 9, due Sep 12 · Cisco advisory updated to v2.5 the same day - **Tags:** `cve` `cisco` `rce` `kev` CISA added Cisco Secure Firewall Management Center CVE-2026-20079 to the KEV catalog on Sep 9 with a federal remediation deadline of Sep 12 — a three-day window. CVSS 10.0 was assigned by Cisco PSIRT as CNA (NVD carries it as a secondary score). The unauthenticated web-interface auth bypass (CWE-288) chains to root RCE "via an improper system process that is created at boot time"; a public PoC by CyberAuth was independently reproduced against FMC 10.0.1-1 with "confirmed uid=0(root) execution" (Full Disclosure, Aug 20). Cisco's v2.5 advisory confirms "the Cisco PSIRT became aware of active exploitation" in August and warns hot fixes "may not address existing compromise," with an IoC check for `/var/tmp/license.tmp`. There are no workarounds. **Why it matters:** a 3-day KEV due date plus the advisory's own hedge — patching does not evict an implant — makes this a forensics problem, not a patch-Tuesday line item. The scorer discipline matters too: 10.0 is Cisco-assigned, not NVD-analyzed. [`🔗 Cisco advisory cisco-sa-onprem-fmc-authbypass-5JPp45V2`](https://sec.cloudapps.cisco.com/security/center/content/CiscoSecurityAdvisory/cisco-sa-onprem-fmc-authbypass-5JPp45V2) · [`🔗 Full Disclosure: PoC reproduction with uid=0`](http://seclists.org/fulldisclosure/2026/Aug/80) --- ## 4. Nous Research's Hermes Agent ships v0.21.1 — a rollup of 5,139 commits, still +4.2k stars/week at ~244k - **Velocity:** ▮▮▮ trending - **Source:** GitHub Trending weekly · 243,806★ (+4,221/wk) · release v0.21.1 Sep 7 - **Tags:** `agent-runtime` `self-improving-agents` `open-source` Nous Research's self-improving terminal agent (skills created from experience, Honcho user modeling, messaging gateways for Telegram/Discord/Slack/WhatsApp/Signal, seven execution backends) shipped v0.21.1 on Sep 7 — an explicit "rollup of main since v0.21.0": 5,139 non-merge commits, 4,364 files, 632 merged PRs, with curated notes deferred to v0.22.0. The README's caveats are concrete: Windows Defender flags the bundled `uv.exe` as a false positive (an attestation-verification procedure is published), and a dev venv inside the checkout "can be wiped by a relative-path command the agent runs against its own checkout." The counterweight: 5k+ open issues and 5k+ open PRs. **Why it matters:** commit volume in a single patch rollup is unusual even for this repo and signals the self-improving-agent category is consolidating around Hermes — but the maintenance load is the part of the story the star count omits. [`🔗 NousResearch/hermes-agent`](https://github.com/NousResearch/hermes-agent) · [`🔗 v0.21.1 release notes`](https://github.com/NousResearch/hermes-agent/releases) --- ## 5. NeoHorse-1 claims a step toward recursive self-improvement — labeled a "prototype" in its own paper - **Velocity:** ▮▮ rising - **Source:** Hugging Face papers · #1 of Sep 9 · 352 submitters / 70 paper upvotes - **Tags:** `self-improvement` `agentic` `post-training` The TokenRhythm-published NeoHorse-1 paper (arXiv, Sep 8, a 36-author "NeoHorse Team") describes agent-native 4B/9B models trained on logs from a routed heterogeneous model pool: routing signals structure a 3-stage SFT curriculum plus routing-guided on-policy distillation, closed by an evaluation-selection-update feedback loop. Macro-average across 11 benchmarks rises 58.94→64.87 (4B) and 65.60→69.04 (9B), with the post-trained 4B largely closing the gap to the 9B base. The paper itself calls the loop "an initial prototype of this feedback-driven process" — recursion across iterations is a stated future path, not a demonstrated result. **Why it matters:** the RSI framing will get quoted without the caveat; the honest headline is "harness-mediated agentic post-training with a feedback-loop design." Per feed discipline, the authors' own disclaimer is the story's anchor, not its footnote. [`🔗 arXiv:2609.08183`](https://arxiv.org/abs/2609.08183) · [`🔗 Hugging Face papers`](https://huggingface.co/papers) --- ## 6. Fortinet CVE-2025-25249: a 30,000-target "PivotC2" campaign teardown — and the CVE enters KEV the next day - **Velocity:** ▮▮ rising - **Source:** SOCRadar research · published Sep 8 · CISA KEV added Sep 9 (due Sep 12) - **Tags:** `fortinet` `firewall` `rat` `kev` SOCRadar's Threat Research Unit published a full teardown of a campaign exploiting CVE-2025-25249 — a heap overflow in the FortiOS/FortiSwitchManager `cw_acd` CAPWAP daemon (UDP 5246) — to deploy "PivotC2," a Node.js post-exploitation RAT whose code carries AI-assisted comments. Attacker files show 30,000+ targeted IPs and 178 confirmed infections (US-heavy, two full intrusions with data exfiltration); SOCRadar assesses with high confidence a Russian-speaking, financially motivated crew, active since at least July. The next day CISA added the CVE to KEV. Scorer note: NVD scores 9.8, Fortinet's own CNA score is 8.1 (`AC:H`). SOCRadar's FAQ answers its own question bluntly: "Does patching remove PivotC2? No." **Why it matters:** a 10-month-old patched CVE with a weaponized chain (ASLR bypass, heap grooming, ROP via FortiOS's own Node.js runtime) is still eating firewalls — and patching closes the exploitation path without removing the implant, so credential rotation on `fsv_sync.dat`-harvested devices is mandatory. Two CNA/NVD scores disagreeing by 1.7 points is the week's scorer lesson. [`🔗 SOCRadar: CVE-2025-25249 PivotC2 teardown`](https://socradar.io/blog/cve-2025-25249-pivotc2-fortigate-rat/) · [`🔗 NVD: CVE-2025-25249`](https://nvd.nist.gov/vuln/detail/CVE-2025-25249) --- ## 7. Desert Ant Labs launches 18 tiny on-device models — free below 100k devices/month - **Velocity:** ▮▮ rising - **Source:** Hacker News · 326+ pts · 85 comments · ~9h ago (~19:39 UTC+8) - **Tags:** `on-device-ai` `edge-inference` `sdk` A new European lab (founded by the maker of the Detail video app) launched Sep 8 with 18 task-specific on-device models (12 stable) behind one Swift/Kotlin/JS SDK, free below 100k monthly active devices: Voz transcribes 10 minutes of audio in 2s on iPhone (claimed 4.7× faster than Whisper, 319× realtime on M3 Ultra vs 78× for Apple SpeechAnalyzer), 9MB Clear for audio enhancement, 12MB Redact for PII masking, 284MB Clips for video. The post's own caveats: Redact catches 88.8% of PII vs GLiNER-PII's 91.1%, all benchmarks are self-reported, and Clear's figure is "best of three." **Why it matters:** a coherent, priced SDK alternative to cloud APIs for the transcription/redaction/classification tier of app development — the free tier, no tokens and no logins attacks per-token pricing directly, and the privacy argument ("never uploaded") does real work. Every performance delta is a vendor benchmark, and the post says so. [`🔗 Desert Ant Labs launch post`](https://desertant.com/blog/introducing-desert-ant-labs/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49624823) --- ## 8. Imbad0202/academic-research-skills — a 4-skill, 32-agent research pipeline at 47k stars, honest about its fabrication blind spot - **Velocity:** ▮▮ rising - **Source:** GitHub Trending weekly · 47,254★ (+2,430/wk) · changelog v3.21.2 Sep 6 - **Tags:** `agent-skills` `research-tools` `claude-code` A four-skill suite covering research → write → review → revise → finalize: a 13-agent deep-research team (8 modes including PRISMA systematic review), a 12-agent paper pipeline (MD/DOCX/LaTeX→PDF), a 7-agent multi-perspective reviewer with a Devil's Advocate role, and a 10-stage orchestrator with "mandatory integrity gates." The README's philosophy is "AI is your copilot, not the pilot," and its caveats are unusually blunt: "a consistently reported fabrication can pass these checks" (it verifies reported content, not whether experiments were run), live reviewer output is `NOT_CALIBRATED`, and it explicitly refuses to be a "humanizer." License is CC BY-NC 4.0 — non-commercial only. **Why it matters:** it shows the skills ecosystem maturing past single-trick repos into full multi-agent pipelines — while the NC license and the fabricated-experiment blind spot are exactly the two things downstream users will trip on. [`🔗 Imbad0202/academic-research-skills`](https://github.com/Imbad0202/academic-research-skills) · [`🔗 GitHub Trending weekly`](https://github.com/trending?since=weekly) --- ## 9. Distillation fingerprint: Qwen3.8 shows the strongest pull toward GPT-5.5 Pro answers under reasoning-prefill test - **Velocity:** ▮▮ rising - **Source:** Hacker News · 90+ pts · 35 comments · ~3h ago (~01:24 UTC+8) - **Tags:** `distillation` `benchmarks` `qwen` Yu Zhang (wsxiaoys) extended his "distillation fingerprint" test: seed a model's reasoning channel with the first 1% of GPT-5.5 Pro's reasoning, then measure n-gram overlap with the teacher's final answer across 45 problems. Qwen3.8 A95B jumped from 16.79% to 34.97% (+18.18pp, gains in all categories), while DeepSeek V4 Flash moved −1.17pp, Inkling +0.46pp, and Kimi K3 +4.54pp. The author's conclusion is deliberately narrow: "Qwen may have learned from GPT-5.5 Pro, or from a closely related GPT model" — and a prior run showed Qwen barely shifted toward Opus 4.8. **Why it matters:** an n-gram overlap test is suggestive, not proof of distillation — 45 problems and a single method must travel with the claim. Distinct from the quantization-benchmark Qwen3.8 story we carried Sep 9: this one is about what the model may have been trained on, not how it quantizes. [`🔗 wsxiaoys' gist: the method and results`](https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49630026) --- ## 10. AuK — an open-weights 1.5B foundation model unifying speech generation and editing - **Velocity:** ▮▮ rising - **Source:** Hugging Face papers · #2 of Sep 9 · 163 submitters / 152 paper upvotes - **Tags:** `speech` `tts` `open-weights` AuK (arXiv 2609.08936) unifies speech generation and instruction-based editing in a ~1.5B model trained on ~3B instruction-audio instances / 1.95M hours, built on Qwen2.5-Omni semantic conditioning, an audio VAE, and a hybrid rectified-flow MMDiT/DiT — with code and weights released. Self-reported numbers: 2.65% avg WER / 0.795 SIM on Seed-TTS-Eval, best on SpeechEditBench content/prosody/acoustic; a distilled AuK-Flash runs 4 steps with a claimed 4.5× speedup. Stated limitations: weak native free-form instruction following (a Prompt Enhancer router is still needed) and Chinese homophone errors that RL can't fix. Attribution note: the HF listing's "Tencent Hunyuan" submitter tag is not confirmed by the paper — the author group is the F5-TTS academic team. **Why it matters:** a genuinely open, small, editing-capable speech foundation model is a direct open-source answer to closed TTS stacks — the metrics are self-reported but the released weights make them checkable. [`🔗 AuK on Hugging Face papers`](https://huggingface.co/papers/2609.08936) · [`🔗 arXiv:2609.08936`](https://arxiv.org/abs/2609.08936) --- ## 11. Tencent Hunyuan's Gander splits full-duplex voice agents into a 9B "cerebellum" and a training-free "brain" - **Velocity:** ▮ steady - **Source:** Hugging Face papers · #3 of Sep 9 · 105 submitters / 137 paper upvotes - **Tags:** `voice-agents` `full-duplex` `architecture` The Omni Interaction Agent Technical Report presents Gander: a 9B streaming Thinker-Talker full-duplex "front cerebellum" (chunk-wise listen/speak/interrupt decisions, no external VAD, ~2-minute sliding context) plus a training-free plug-and-play task agent as the "back brain." On Full-Duplex-Bench v3 it leads on turn-taking (100.0) and interruption latency (8.0 vs 13.5 for GPT-Realtime) but trails commercial systems on task accuracy (Pass@1 0.400 vs 0.600 best). Its own limitations: a 51.6% filler rate, a 6.08-point WorldSense regression vs its base model, ASR-text-only brain-cerebellum coupling, and no post-training/RL yet. **Why it matters:** the split architecture lets reasoning upgrades ride without retraining the interaction layer — a pragmatic pattern likely to be copied — but the paper is unusually candid that task accuracy still trails GPT-Realtime. [`🔗 arXiv:2609.08977`](https://arxiv.org/abs/2609.08977) · [`🔗 Hugging Face papers`](https://huggingface.co/papers) --- ## 12. OpenWAM — a modular world-action-model pretraining stack claims #1 on a real bimanual robot leaderboard, doubling pi0.5 - **Velocity:** ▮ steady - **Source:** Hugging Face papers · 48+ submitter upvotes · GitHub 344★ - **Tags:** `world-model` `robotics` `open-source` OpenWAM decomposes world-action-model pretraining into Infra (composable modules, 8 sim benchmarks), Study (controlled experiments on knowledge inheritance and world-action synergy), and OpenWAM-α, pretrained on ~6,400 hours (~518.5M frames) of egocentric human + robot data. Claimed results: SOTA on the mobile bimanual EBench and #1 on real-world RoboDojo-Real, "doubling pi0.5's success rate." The claimed edge is specifically out-of-domain generalization, and all results are author-reported. **Why it matters:** world-action models are where world-model research and VLA robotics converge; a fully open stack (code, weights, data recipes) makes the approach reproducible in a way prior closed WAM work was not. [`🔗 OpenWAM on Hugging Face papers`](https://huggingface.co/papers/2609.07398) · [`🔗 arXiv:2609.07398`](https://arxiv.org/abs/2609.07398) --- ## 13. Procedural Graphs — a Google-led paper gives agents self-evolving "what-to-do" memory - **Velocity:** ▮ steady - **Source:** arXiv + Hacker News · 24+ pts · ~3h ago (~01:13 UTC+8) - **Tags:** `agent-memory` `research` `planning` arXiv:2609.09153 (submitted Sep 8; Yuxing Lu, Yicheng Chen, Shanchan Wu, Sercan Ö. Arık) introduces Procedural Graphs — "(procedure, relation, procedure)" triplets as the procedural analog of knowledge graphs. A guidance model turns the local subgraph into step-level hints that "bias the solver's next action without dictating it," and an LLM refiner edits graph topology from failed-vs-successful trajectory contrasts, keeping only edits that hold on validation. Claims: evolved graphs match or surpass hand-designed ones, can repair a flawed expert prior, and beat memory-based baselines across datasets and LLMs. No limitations section is visible in the abstract — benchmark specifics live in the 36-page body. **Why it matters:** agent memory has been mostly episodic/factual; a self-evolving procedural layer that explicitly learns tool-ordering and preconditions is a different primitive — and it lands directly in the MCP/skills tooling conversation. [`🔗 arXiv:2609.09153`](https://arxiv.org/abs/2609.09153) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49629868) --- ## 14. Geiger — `npx geiger-scan` inventories every agent, MCP server, and skill on your machine - **Velocity:** ▮ steady - **Source:** Show HN · 33+ pts · 19 comments · ~6h ago (~22:54 UTC+8) - **Tags:** `agent-security` `mcp` `supply-chain` Atomburst's Geiger (57★, JavaScript, MIT, zero runtime dependencies) is a read-only scanner that reads known config locations — Claude Code MCP/hooks/plugins/skills/subagents, MCP hosts (Cursor, Windsurf, VS Code, Cline, Zed…), agent CLIs (Codex, Gemini CLI, Aider, Goose…) — without executing npm, and labels each finding EXECUTES / HOLDS-SECRETS / BROAD-FILESYSTEM / NETWORK plus an origin class (UNKNOWN-ORIGIN included); a `--diff` baseline mode acts as a drift alarm for CI/cron. Honest limits from the README: it "reads configuration, not runtime behavior," misses agents in containers/WSL/other user accounts, and "origin ≠ trustworthiness." **Why it matters:** with skills and plugins being installed from GitHub at trending scale, a machine-level inventory/audit tool answers the obvious next question: what did I actually install and what can it reach? [`🔗 Atomburstofficial/geiger`](https://github.com/Atomburstofficial/geiger) · [`🔗 Show HN discussion`](https://news.ycombinator.com/item?id=49627646) --- ## 15. Red Hat Hawtio Operator signing oracle (CVE-2026-78234, CVSS 9.9) — cross-tenant impersonation to RCE on OpenShift - **Velocity:** ▮ steady - **Source:** Red Hat advisory · surfaced Sep 8–9 · SecurityOnline writeup Sep 9 - **Tags:** `kubernetes` `openshift` `cve` The hawtio-operator (Red Hat build of Apache Camel tooling) reads the OpenShift Service CA private signing key and mints client certificates with an attacker-chosen Common Name — effectively "a signing oracle to any namespace editor," enabling cross-tenant service impersonation and, via Jolokia MBean invocation, RCE. CVSS 9.9 was assigned by Red Hat as CNA but is explicitly "preliminary and subject to review," and Red Hat rates the impact only *Important* because authentication (edit rights in any namespace) is required. A companion advisory, CVE-2026-77968 (CVSS 8.2, NVD-published Sep 8), covers the operator's over-broad cluster-wide Secret read — the two are distinct CVEs, not one bug. **Why it matters:** exploitation requires only cluster edit-role access — a common grant — and flips it into cluster-wide certificate forgery. Mitigations are configuration changes (CSR API, RBAC tightening, cert rotation), not just an upgrade; no confirmed in-the-wild exploitation and no public PoC so far. [`🔗 Red Hat: CVE-2026-78234`](https://access.redhat.com/security/cve/CVE-2026-78234) · [`🔗 SecurityOnline analysis`](https://securityonline.info/hawtio-operator-vulnerability-cve-2026-78234/) --- ## 16. petergyang/no-ai-slop — the anti-AI-voice skill race continues: 20+ slop patterns, 7.8k stars in days - **Velocity:** ▮ steady - **Source:** GitHub Trending weekly · 7,792★ (+1,038/wk) · #20 weekly - **Tags:** `agent-skills` `writing-tools` `llm` An install-as-a-skill linter for AI tells ("It's not X. It's Y." binary contrasts, throat-clearing openers, "a testament to" puffery, fake-profound endings…), runnable as `/no-ai-slop` in Claude Code/Codex/ChatGPT or via `npx skills add`. Detection mode deliberately flags style "without guessing whether AI wrote the text," and the README concedes the core tension itself: AI editing tends to "smooth away" personal quirks — the exact risk the skill exists to mitigate. Only 10 of the claimed 20+ patterns are enumerated publicly (the rest live in `SKILL.md`), there are no releases, and it cannot run standalone. **Why it matters:** the second wave of the humanizer moment — evidence the "agent skill as one-file product" channel now carries writing tools to thousands of stars in a week, and also that these repos ship as undocumented rule files. The pattern taxonomy is English-LLM-specific; each language needs its own tell list. [`🔗 petergyang/no-ai-slop`](https://github.com/petergyang/no-ai-slop) · [`🔗 GitHub Trending weekly`](https://github.com/trending?since=weekly) --- ## 17. GNU Radio runs entirely in the browser — WASM port with live SDR over WebUSB - **Velocity:** ▮ steady - **Source:** Hacker News · 126+ pts · 16 comments · ~5h ago (~23:53 UTC+8) - **Tags:** `sdr` `webassembly` `gnu-radio` gnuradioworld.com is a GNU Radio Companion-style flowgraph editor and runtime where the DSP stack and Qt GUI sinks are compiled to WebAssembly; it reads/writes native `.grc` files, plots live spectrum/waterfall/constellation, and talks to RTL-SDR, PlutoSDR, or HackRF over WebUSB — "no Python, no install, no server." Author Marc Lichtman (777arc). The README's honest numbers: WASM is measurably slower than native (12.1 vs 24 Msps on a decimating FIR), there's no general Python runtime (Python-only blocks need C++ ports), and it requires SharedArrayBuffer/COOP-COEP hosting on Chrome or Firefox. **Why it matters:** it removes the install barrier that has always gated SDR experimentation — the GNU Radio module ecosystem (gr-satellites, gr-adsb, gr-lora_sdr…) becomes a URL, and the port publishes its own performance losses rather than hiding them. [`🔗 gnuradioworld.com`](https://gnuradioworld.com/) · [`🔗 777arc/gnuradio-world`](https://github.com/777arc/gnuradio-world) --- ## 18. Read the Docs post-mortem: a 10-day, 5.5M-req/min DDoS designed to inflate the bill - **Velocity:** ▮ steady - **Source:** Read the Docs blog (Sep 8) · HN 98+ pts · ~5h ago (~23:55 UTC+8) - **Tags:** `ddos` `infrastructure` `cloudflare` Read the Docs disclosed a nearly ten-day June 2026 attack that peaked above 5.5M requests/minute (~100× normal, 10× larger than any prior incident) from millions of IPs across hundreds of ASNs. The attack randomized HTTP/TLS fingerprints to beat JA3/JA4 fingerprinting, targeted only cache-miss URLs (unique 404s, uncached 302s), and used a "yo-yo pattern" to maximize auto-scaling costs. Mitigation: caching 404s/redirects at the edge, bot-score-plus-per-IP rate limits, and a penalty-box rule system managed via Terraform. The hedges are the value: two defenses failed outright (protocol-inconsistency checks and JA4 both defeated), they declined Cloudflare's Under Attack Mode ("would break every API integration"), and they call the calm "an extended reprieve," with residual attack traffic continuing at publication. **Why it matters:** a template post-mortem for anyone running public docs infrastructure — IP blocking is "obsolete for distributed attacks," and the modern DDoS goal is your cloud bill. The post is Sep 8; the attack was June — don't read this as an ongoing incident. [`🔗 Read the Docs: the 2026 DDoS attack`](https://about.readthedocs.com/blog/2026/09/2026-ddos-attack/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49628614) --- ## 19. Google Ads flags a signed, notarized macOS app as "malicious software" — reinstated only after HN attention - **Velocity:** ▮ steady - **Source:** xlii.space (Sep 9) · HN 306+ pts · 185 comments · ~9h ago (~19:43 UTC+8) - **Tags:** `google-ads` `app-distribution` `false-positive` The developer of RACE, a signed-and-notarized native macOS terminal multiplexer, spent $500 on a first Google Ads campaign and had the account suspended for "Malicious software" and "Compromised Site" — with zero detail on what triggered either flag. He documents exhaustive clean checks (Safe Browsing, Search Console, VirusTotal with hash, notarization, JS bundle review), four auto-rejected appeals, and a week-long block. A top-of-post edit after the HN thread: "Through the apparent magic of Hacker News, my Google Ads account has been reinstated… Still no explanation of what triggered the suspension." **Why it matters:** a concrete case of opaque automated security-flagging at an ad-platform chokepoint, where the only recourse that worked was viral visibility — and reinstatement correlated with attention, not with any evidence change. The author's own caveat holds: Safe Browsing clearance "does not establish what Google Ads detected." [`🔗 xlii.space: Malicious software on Google Ads`](https://xlii.space/eng/malicious-software-on-google-ads/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49624856) --- ## 20. RadixArk Miles v0.1 lands with a tech report — since we covered it Sep 4, the enterprise RL fork has shipped - **Velocity:** ▮ steady - **Source:** arXiv tech report (Sep 8) · Hugging Face papers · GitHub 2.7k★ - **Tags:** `rl` `post-training` `open-source` **Update:** when we covered radixark/miles on Sep 4, it was "an enterprise fork of slime emerging for large-scale RL post-training." It has since shipped Miles v0.1 (Apache-2.0) with a 34-page tech report (arXiv Sep 8): SGLang rollout engines, Megatron-LM or FSDP trainer backends, three weight-sync transports, plus LoRA RL, on-policy distillation, and diffusion-model support. The report's case study runs fully asynchronous agentic RL on GLM-5.2 744B-A40B for terminal-use coding on 64 GB300 GPUs, hitting a 263s median step time; the README confirms low-precision training (MXFP8/NVFP4), TITO, MoE routing replay, and fault tolerance, with day-0 support for GLM-5.2/DeepSeek-V4/Kimi-K3. **Why it matters:** frontier-scale RL infrastructure is the scarcest layer of the open stack; a validated, enterprise-framed entrant with published case-study numbers signals post-training tooling is commoditizing. The GLM-5.2 numbers are the vendor's own report — one source, not an independent benchmark. [`🔗 arXiv:2609.08368 tech report`](https://arxiv.org/abs/2609.08368) · [`🔗 radixark/miles`](https://github.com/radixark/miles) --- ## 21. DeepSeek V4.1 Flash launches openly — since we covered the internal beta Sep 9, the weights, tech report and MIT license have landed - **Velocity:** ▮▮▮ trending - **Source:** Hugging Face + Hacker News · 480+ pts · 260 comments · ~5h ago (~15:00 UTC+8) - **Tags:** `deepseek` `open-weights` `moe` **Update:** on Sep 9 we covered V4.1 Flash as "an internal beta with a two-day window." The public launch has now happened: DeepSeek-V4.1-Flash is on Hugging Face under MIT, code and weights, with a tech report. The model page: a 552B-backbone MoE (485B stored + a 196B sparsely-accessed "Engram" conditional-memory module) with only ~8B active at prefill / ~16B at decode, a 40-layer Causal Encoder-Decoder, FP4 KV caching plus CSA2 sparse attention claiming "890 bytes per token" of global KV cache (~¼ of V4-Flash), 384 routed experts (6 active), a ViT vision encoder, and a 45T-token multimodal pretraining corpus with 1M context. Instruct numbers at max reasoning effort: GPQA Diamond 90.9, Codeforces 3471, Terminal-Bench 2.1 90.6, HLE 36.8. Honest friction in the release itself: no Jinja chat template (a Python reference and a Rust toolkit ship instead), and the model card notes DeepSWE ranges 65.6–74.2 depending on agent scaffold. **Why it matters:** the beta's promised architecture is now inspectable — and the model card's own scaffold-sensitivity note is the caveat that belongs in every benchmark quote of it. The HN thread's practical floor: at ~552B total it no longer fits the "flash" niche for local users (one estimate puts q4 just short of 256 GB). [`🔗 deepseek-ai/DeepSeek-V4.1-Flash`](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49639090) --- ## 22. "ShieldCrash" — a third bypass in the Microsoft Defender saga drops the day after Patch Tuesday, again with a public PoC - **Velocity:** ▮▮▮ trending - **Source:** BleepingComputer (Sep 9) + GitHub PoC · 228★ · ~13h ago (~07:00 UTC+8) - **Tags:** `windows` `defender` `lpe` `zero-day` An anonymous researcher ("Nightmare Eclipse" / MSNightmare) published ShieldCrash on Sep 9, immediately after September Patch Tuesday: a claimed bypass of Microsoft's fix for ShieldBreak (CVE-2026-69414), which was itself a bypass of June's RoguePlanet Defender flaw. The PoC triggers "an arbitrary file read as SYSTEM" on fully patched Windows 10/11/Server, with the README claiming "all supported windows versions are affected." The caveats are the researcher's own: it is a "skeleton PoC" for file *read* only — "might rework this later into a full SYSTEM PoC." There is no CVE for ShieldCrash, no vendor confirmation, no evidence of in-the-wild exploitation, and Microsoft has not responded. It continues an ongoing disclosure dispute; Microsoft has previously warned of legal action, and several of the researcher's earlier findings (LegacyHive, BlueHammer, RedSun, UnDefend) remain unpatched. **Why it matters:** the RoguePlanet→ShieldBreak→ShieldCrash chain is now a three-round patch-bypass series against the same component — a case study in how unsatisfied an adversary can keep a patched bug, and in adversarial disclosure running ahead of both bounty process and vendor response. [`🔗 BleepingComputer: ShieldCrash zero-day`](https://www.bleepingcomputer.com/news/security/new-microsoft-defender-shieldcrash-zero-day-grants-system-access/) · [`🔗 MSNightmare/ShieldCrash PoC`](https://github.com/MSNightmare/ShieldCrash) --- ## 23. Apple unveils iPhone Duo — its first foldable, at ~$2,000, with the crease debate baked into launch day - **Velocity:** ▮▮▮ trending - **Source:** Apple + Hacker News · 1,240+ pts · 2,176 comments · ~14h ago (~06:00 UTC+8) - **Tags:** `apple` `hardware` `foldable` Apple's keynote unveiled iPhone Duo, its first book-style foldable: a square closed form factor with a 5.4-inch outer display (the discontinued iPhone mini's size), heavy crease-minimization messaging, and an October 23 launch at roughly $2,000. The launch-day HN thread (2,176 comments) already carries the counter-evidence: multiple photo/video reports of a visible crease on dark backgrounds with the screen off, skepticism about long-term hinge durability, and notes that it drops a lens option vs the iPhone 18 Pro. The keynote also positioned a leadership handoff from Tim Cook to John Ternus, alongside AirPods 5, Apple Watch Series 12, and iPhone 18 Pro announcements. **Why it matters:** the largest HN thread of the day is a hardware launch whose contested claim — the crease — is exactly the kind of marketing-vs-reality gap that gets settled by owners, not keynotes; the $2,000 price makes early-adopter risk real. For developers the 5.4-inch outer/square inner split is a new layout target. [`🔗 Apple: iPhone Duo`](https://www.apple.com/iphone-duo/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49630931) --- ## 24. Raschka on GPT-6 Astra and "looped transformers" — weight reuse, not a hidden-CoT conspiracy - **Velocity:** ▮▮ rising - **Source:** Sebastian Raschka + Hacker News · 452+ pts · 146 comments · ~16h ago (~04:00 UTC+8) - **Tags:** `transformers` `interpretability` `reasoning` Reacting to reporting that framed GPT-6 Astra's rumored looped-transformer architecture as a secret technique for hiding reasoning, Raschka's piece argues looping (reusing weights across stacked layers) is primarily a GPU-memory-saving parameter-sharing trick — "still just producing one token at a time" — not inherently a CoT-monitoring hazard. The HN thread sharpens both sides: commenters note that if loop depth is chosen dynamically per token, the computation becomes far richer, and point to the Astra system card's own table showing a trivia answer solved while the visible CoT discusses something unrelated — unusually high "CoT controllability." Will Merrill's work on how much CoT different problems require gets cited as the right theoretical frame. **Why it matters:** since we covered the GPT-6 Astra launch on Sep 4, this is the first serious architecture-level critique of its "hidden reasoning" narrative — and the useful conclusion is that latent-space iteration is a real interpretability question that doesn't need a leak-based framing to matter. [`🔗 Raschka: GPT-6 Astra, looped transformers, and hidden reasoning`](https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49627370) --- ## 25. bilawalsidhu/gods-eye-view — a browser "spy satellite simulator" with only real data re-trends at +1,588/day - **Velocity:** ▮▮ rising - **Source:** GitHub Trending · 22,200★ (+1,588 today) · MIT - **Tags:** `osint` `cesium` `data-visualization` A CesiumJS + Google Photorealistic 3D Tiles globe that layers live public feeds onto a "spy satellite" interface: 11,000+ aircraft (OpenSky/adsb.lol), ships (AISStream), an 838-object satellite catalog (CelesTrak), earthquakes (USGS), ~800 public CCTV cameras, NASA fire detections, radio stations — 11 of 13 layers work keyless, with NVG/FLIR sensor styles, a military HUD, and voice control via OpenAI Realtime (28 tools, $5 hard session cap). The README's honesty is unusually thorough: "Data may be delayed, incomplete, modeled, inferred, or wrong"; traffic is simulated on real roads; CCTV poses are user-calibrated priors; launch replays are labeled "RECONSTRUCTED ESTIMATE"; and it explicitly refuses to build person-tracking or face recognition. First trended #1 in August; the current wave rides a viral YouTube series. **Why it matters:** the same open feeds that power OSINT tooling assembled into one cinematic interface — with the limits stated in-product rather than discovered by users. The trigger is the YouTube wave, not new code (24 commits total), which matters for anyone reading the star count as momentum. [`🔗 bilawalsidhu/gods-eye-view`](https://github.com/bilawalsidhu/gods-eye-view) · [`🔗 GitHub Trending`](https://github.com/trending) --- ## 26. CISA: WatchGuard Firebox iked RCE (CVE-2025-14733, 9.3) is now feeding ransomware attacks — 9,000 boxes still unpatched - **Velocity:** ▮▮ rising - **Source:** CISA + BleepingComputer (Sep 10) · WatchGuard PSIRT advisory updated Aug 10 - **Tags:** `cve` `firewall` `ransomware` `kev` CISA said this week that CVE-2025-14733 — an out-of-bounds write in the Firebox `iked` IKEv2-VPN handler, unauthenticated, CVSS 9.3 (vendor-assigned), KEV-listed since December 2025 — is now known to be used by ransomware gangs, without naming groups or sharing details. WatchGuard patched in December 2025 and confirmed in-the-wild exploitation then; Shadowserver counted 115,000+ exposed unpatched Fireboxes at the time, and roughly 9,000 remain vulnerable nine months later. Two hedges matter: exploitation requires an IKEv2-VPN configuration, and even devices where that config was deleted may still be exposed if a branch-office VPN to a static peer remains. **Why it matters:** a nine-month-old patch with a shrinking-but-persistent unpatched population is exactly how ransomware crews save-target — and the "we deleted the config" mitigation being incomplete is the detail most likely to be wrong in real inventories. WatchGuard's SME footprint (250,000+ businesses via 17,000+ resellers) makes the long tail long. [`🔗 BleepingComputer: CISA on WatchGuard ransomware attacks`](https://www.bleepingcomputer.com/news/security/cisa-watchguard-rce-flaw-now-exploited-in-ransomware-attacks/) · [`🔗 WatchGuard PSIRT: CVE-2025-14733`](https://psirt.watchguard.com/) --- ## 27. JustVugg/colibri — pure-C inference that streams frontier MoE experts from disk, no "flash" pretense - **Velocity:** ▮▮ rising - **Source:** GitHub Trending · 27,285★ (+157 today) · Apache-2.0 - **Tags:** `inference` `moe` `ssd` Colibrì runs 744B–2.8T MoE models (GLM-5.2, Kimi K3, Inkling, DeepSeek V4 Flash…) on consumer hardware by treating VRAM/RAM/NVMe as one hierarchy: dense parts stay resident (~9.9 GB int4 for GLM-5.2), 19,456 routed experts (~370 GB) stream from disk on demand — "a JIT, but for weights," with routing 71.6% predictable one layer ahead. Published numbers are honest about the floor: 5.8–6.8 tok/s on 6× RTX 5090, ~1.8 tok/s warm on a 128 GB CPU-only box, 0.05–0.1 tok/s cold on 25 GB. The README lists what's unproven (placement, SSD striping, auto-planning) and reports speculative decoding as a measured net loss in its own tests (MTP: −32% near 85% expert hit; DeepSeek drafters default off). **Why it matters:** distinct from the Kimi-K3-on-MacBook story we carried Sep 9 (a one-off four-SSD stunt), colibri is a maintained open engine for the same trick — and its decision to publish failure modes (quantization container rules, drive-dependent `O_DIRECT`) alongside tok/s is what makes the numbers usable. [`🔗 JustVugg/colibri`](https://github.com/JustVugg/colibri) · [`🔗 GitHub Trending`](https://github.com/trending) --- ## 28. little-lm: a 3.8B LLM trained from scratch to 0.384 CORE for $998 — with the measurement caveats printed - **Velocity:** ▮▮ rising - **Source:** Hacker News · 91+ pts · 14 comments · ~4h ago (~16:00 UTC+8) - **Tags:** `pretraining` `nanochat` `reproducibility` Hugo Vergnes (video understanding at Apple, evenings project) trained a 3.8B Llama-style model from random weights: 65.3B tokens of ClimbMix, 43 hours on 8× rented B200s, $998, 0.384 CORE — vs nanochat d32's 0.310 at ~$1,000 and GPT-2 1.5B's 0.2565. The write-up's value is the negative space: FineWeb-Edu was rejected after a failed early run; the 1024→2048 context rerun that produced the headline score was largely a *measurement* fix (SQuAD and BoolQ prompts hadn't fit in 1024 tokens — those two tasks were ~83% of the gain, the other 19 tasks moved +0.008); CORE moved ~7× more than loss, which he flags as a caution for anyone using CORE to make decisions. FP8 GEMMs, Muon+AdamW, trapezoidal LR, ResFormer-style value embeddings (+3.2% CORE for 721M params) are documented with what worked and what didn't. **Why it matters:** "frontier-adjacent pretraining at hobbyist budget" is becoming a reproducible genre — and this entry is worth reading specifically because it shows how much of a benchmark jump can be harness artifact rather than learning. [`🔗 little-lm 3.8B writeup`](https://hugovergnes.github.io/little-lm-3-8b/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49637435) --- ## 29. AlexsJones/llmfit — one command to answer "which models actually run on this machine?" - **Velocity:** ▮ steady - **Source:** GitHub Trending · 35,510★ (+247 today) · MIT · Rust - **Tags:** `local-llm` `hardware` `developer-tools` llmfit detects CPU/RAM/GPU/VRAM (CUDA, Apple Silicon, ROCm, oneAPI; multi-GPU and MoE-aware) and scores catalog models on four axes — memory fit, estimated speed, quality, context — with quantization awareness (GGUF, AWQ, GPTQ, EXL2), a TUI, a web dashboard, and REST endpoints. Its honesty lives in `llmfit info`: speed and memory figures are model-based estimates grounded in a memory-bandwidth model and community `bench --share` submissions, and accuracy depends on community data matching your hardware — real measurements replace estimates locally. **Why it matters:** the "will it run" question is answered daily by trial-and-error across quantization forums; a tool that centralizes it — and keeps its estimate-vs-measured distinction visible — is quiet but broadly useful infrastructure for the local-model wave. [`🔗 AlexsJones/llmfit`](https://github.com/AlexsJones/llmfit) · [`🔗 GitHub Trending`](https://github.com/trending) --- ## 30. diegosouzapw/OmniRoute — a self-hosted AI gateway claiming ~1.47B free tokens/month by stacking provider free tiers - **Velocity:** ▮ steady - **Source:** GitHub Trending · 63,817★ (+591 today) · MIT - **Tags:** `ai-gateway` `routing` `self-hosted` One local OpenAI-compatible endpoint routing to 352 registered providers (152 flagged free) with quota-aware fallback, circuit breakers, key cooldown, 19 "combo" routing strategies, an MCP server (110 tools), and 42-language localization. The README carries its own asterisks: the 1.47B free-tokens/month figure is a best-case aggregate that "move[s] both ways" as providers change terms, the ~3B/mo "Radar ceiling is not a guarantee," provider counts differ across sections by design (352/356/444), savings percentages are self-reported, and affiliate links are disclosed. **Why it matters:** the free-tier-aggregation gateway is a category that lives and dies on provider goodwill — OmniRoute's usefulness is real (fallback, quota-awareness, one endpoint) but its headline number is exactly the kind that expires; read the caveats before quoting it. [`🔗 diegosouzapw/OmniRoute`](https://github.com/diegosouzapw/OmniRoute) · [`🔗 GitHub Trending`](https://github.com/trending) --- ## 31. Stockfish 19 released — up to +44 Elo, a new SFNNv16 net, and universal binaries - **Velocity:** ▮ steady - **Source:** stockfishchess.org (Sep 5) · HN 83+ pts · GPL - **Tags:** `chess` `nnue` `open-source` Stockfish 19 (Sep 5) gains up to 44 Elo head-to-head over Stockfish 18 and keeps the engine championship lead. The engineering changes are the story: a new SFNNv16 NNUE network trained with quantization-aware training on hundreds of billions of positions rescored by a strong Leela net (dropping redundant threat features, adding pawn-pair features, retiring the secondary net), universal binaries that auto-detect CPU capabilities, native RISC-V (RVV) and LoongArch support, WebAssembly targets, and strict invalid-input termination. **Why it matters:** chess engines remain the longest-running open benchmark community in existence; QAT-trained NNUE and universal binaries are the kind of unglamorous engineering that keeps a 15-year-old project both dominant and installable by non-specialists. [`🔗 Stockfish 19 announcement`](https://stockfishchess.org/blog/2026/stockfish-19/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49599992) --- ## 32. Automattic's board forces Matt Mullenweg into leave of absence — WordPress.org says the open-source project is unaffected - **Velocity:** ▮ steady - **Source:** TechCrunch (Sep 9) + Hacker News · 325+ pts · 215 comments · ~7h ago (~13:00 UTC+8) - **Tags:** `wordpress` `open-source` `governance` Automattic's board voted Sep 9 to place founder/CEO Matt Mullenweg on paid leave — reportedly against his will, with the resolution delivered 50 minutes before the vote, per his Slack message. CFO Mark Davies is interim CEO; no official reason was given. The backdrop is the WP Engine lawsuit (Automattic's 8% royalty demand, defamation and abuse-of-power claims), the 2024 ultimatum that saw 159 employees quit, and a 16% layoff. WordPress.org's executive director moved quickly to separate the project from the company: "Matt remains the leader of the WordPress project," and priorities continue unchanged. **Why it matters:** the WP Engine dispute turned WordPress governance into a single-founder risk story; the board's move is the first structural check on that concentration. The load-bearing claim for the ecosystem is the project/company separation — which is exactly the promise to watch, given Mullenweg retains control of the .org project. [`🔗 TechCrunch: Mullenweg forced into leave of absence`](https://techcrunch.com/2026/09/09/automattics-board-forces-ceo-matt-mullenweg-into-leave-of-absence/) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49636283) --- ## 33. liquidslr/system-design-notes — 28 chapters of free Alex Xu notes, +1,397 today, and no license at all - **Velocity:** ▮ steady - **Source:** GitHub Trending · 18,422★ (+1,397 today) · no license file - **Tags:** `system-design` `interview-prep` `notes` Chapter-by-chapter notes on Alex Xu's *System Design Interview* Vol 1 & 2 (28 folders, rate limiters through stock exchanges), with per-topic curated external links (Dynamo paper, Discord/Slack engineering blogs, a Stanford consistent-hashing lecture). The caveats are structural: the README says "work in progress," there are only 35 commits, notes are mirrored on a commercial site (pagefy.io), and the repo has **no license file** — default copyright applies, so reuse/redistribution rights are unclear for a derivative of a commercial book. **Why it matters:** the fastest-rising repo of the day is study notes, not software — and its missing license plus the commercial mirror make it a live case of the "trending ≠ yours to reuse" trap, in a week where agent skills built from copyrighted material keep trending too. [`🔗 liquidslr/system-design-notes`](https://github.com/liquidslr/system-design-notes) · [`🔗 GitHub Trending`](https://github.com/trending) --- ## 34. Samsung unveils zHBM prototype — memory stacked directly on the AI accelerator, claiming 8× HBM5 throughput - **Velocity:** ▮ steady - **Source:** THE ELEC + Hacker News · 30+ pts · ~5h ago (~15:00 UTC+8) - **Tags:** `hbm` `memory` `hardware` Samsung disclosed a zHBM prototype that stacks memory directly onto AI accelerator dies instead of alongside them, claiming (all vendor figures, prototype-stage): up to 8× the data-processing performance of 8th-gen HBM (HBM5), 3× better performance-per-watt, and thermal resistance cut by more than half. The HN thread's open question is the right one — where the memory controller lives (a middle stacking layer or the main die) — and there are no production-timeline, capacity, or pricing details. **Why it matters:** memory bandwidth and capacity are the binding constraint on both frontier training and local inference (see colibri, #27); stacking-on-die is the most direct possible attack on that constraint. All numbers are Samsung's own and prototype-stage — treat as a direction, not a spec. [`🔗 THE ELEC: Samsung zHBM prototype`](https://www.thelec.net/news/articleView.html?idxno=12835) · [`🔗 Hacker News discussion`](https://news.ycombinator.com/item?id=49593896) --- ## 35. Show HN: evaluate polynomials with provably fewer multiplications — with a full Lean proof - **Velocity:** ▮ steady - **Source:** Show HN · 100+ pts · 34 comments · ~23h ago (~21:00 UTC+8 Sep 9) - **Tags:** `algorithms` `formal-verification` `lean` Thomas Ahle published a construction for evaluating a pre-known univariate polynomial with fewer multiplications (additions and squarings are cheap; multiplications — especially over finite fields — are not), improving on Horner/Estrin/Knuth-Eve/Pan/Rabin–Winograd lines, plus "an injective polynomial construction for universal hashing that uses N multiplications to hash 2N values with a single random key," improving on Daniel J. Bernstein's. The ~100-page proof is fully machine-verified in Lean, with an interactive site visualizing the circuits. The thread's limits are explicit: rational coefficients blow up (finite fields are the sweet spot), for floating point "use Estrin instead" (FMA, pipelining, stability), it's univariate only, and WyHash/xxh3 aren't polynomial — though the paper shows those heuristic hashes collide far more on adversarial inputs. **Why it matters:** preprocessing-once/evaluate-many polynomial schemes sit under hash maps, MACs and libm — and the rare combination here is a new multiplication-count bound *and* a Lean proof, so the claim is checkable rather than benchmarked. [`🔗 thomasahle.com/fast-polynomials`](https://thomasahle.com/fast-polynomials/) · [`🔗 thomasahle/fast-polynomials`](https://github.com/thomasahle/fast-polynomials) · [`🔗 Show HN discussion`](https://news.ycombinator.com/item?id=49623398) --- ## Metadata | Field | Value | |-------|-------| | Generated | 2026-09-10T20:05:00+08:00 | | Items | 35 | | Sources tracked | 38 (Hacker News, GitHub Trending, Hugging Face papers, arXiv, CISA KEV, NVD, Cisco PSIRT, Full Disclosure, SOCRadar, Red Hat, SecurityOnline, Tailwind CSS blog, Desert Ant Labs, xlii.space, Read the Docs, opusfived.dev, NousResearch, Atomburst, wsxiaoys gist, gnuradioworld.com, 777arc/gnuradio-world, Imbad0202, petergyang, radixark, TechCrunch/CNBC, Fortinet FG-IR-25-084 via NVD, seclists.org, GitHub Trending weekly, BleepingComputer, WatchGuard PSIRT, MSNightmare, apple.com, Raschka/mysterious substack, stockfishchess.org, hugovergnes.github.io, thelec.net, thomasahle.com) | | Update schedule | 04:03, 12:03, 20:03 UTC+8 (3x daily) | | Ranking | Velocity-weighted (recency × engagement acceleration × source authority) | | License | [CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/) | --- [Previous day](../2026-09-09/) · [Raw .md](../2026-09-10.md) · [Archive](../../archive/)