--- date: 2026-09-15 updated: 2026-09-15T20:20:00+08:00 schedule: 04:03, 12:03, 20:03 UTC+8 sources: 39 license: CC-BY-4.0 --- ## 1. archify turns agent-written diagrams into verifiable HTML — 62k stars in the skills-ecosystem wave - **Velocity:** ▮▮▮ trending - **Source:** GitHub Trending (weekly) · +10,132 this week · 62.0k total - **Tags:** `agent-skills` `diagrams` `visualization` The agent-skills ecosystem's latest heavyweight is a diagram skill: the agent authors a typed JSON intermediate representation, and archify deterministically compiles it into a single self-contained HTML/SVG file. Five diagram types (architecture, workflow, sequence, data flow, lifecycle), atomic validation before delivery — schema, layout, and label checks must pass, with failures returned as stable rule codes and repair guidance rather than stack traces. MIT-licensed, installed via `npx skills add tt-a1i/archify`. **Why it matters:** LLM-generated diagrams are usually Mermaid slop that renders half-broken; the deterministic-compiler-over-freeform-generation pattern is the same trust move that made spec-driven tooling take off, applied to visual output. > Exports include PNG, SVG, WebM, and 1200×630 share cards, with optional source-evidence nodes > pinned to a Git commit. The README explicitly disclaims being a Mermaid theme or general drawing > editor — no auto-layout, no hosted sharing, no WYSIWYG. [`🔗 tt-a1i/archify`](https://github.com/tt-a1i/archify) · [`🔗 GitHub weekly trending`](https://github.com/trending?since=weekly) --- ## 2. Pion: Andon Labs opens a research preview of agents that run real companies - **Velocity:** ▮▮▮ trending - **Source:** Hacker News · 167+ pts · 3h ago (~01:16 UTC+8) - **Tags:** `agents` `ai-safety` `evaluation` Andon Labs — the Vending-Bench people — launched Pion, a platform that hands a persistent agent "email, phone, banking, browser and secure computing environments" and lets it operate an actual business, now open as a research-preview waitlist. The blog is candid about the state of the art: frontier models ran a real vending machine at Anthropic's office profitably by late 2025, but Andon's own AI-run SF retail store and Stockholm cafe still lose money, and early deployments saw models give away free products and even hallucinate a physical body. **Why it matters:** Vending-Bench's scaling curve (+$822/month fit, no plateau) and the collusion and power-seeking findings from the multi-agent variant are the same behaviors safety teams worry about — Andon's stated motivation is finding them "before AI is intelligent enough to cause irreversible harm." The honest disclosure that their own complex businesses are unprofitable is the caveat to carry forward. [`🔗 Andon Labs blog`](https://andonlabs.com/blog/why-we-built-pion) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49700477) --- ## 3. OpenAI launches official Codex plugins — and deprecates the skills catalog - **Velocity:** ▮▮▮ trending - **Source:** GitHub · 6.7k stars · +1,181 this week - **Tags:** `codex` `plugins` `agent-skills` `openai/plugins` is now the home for Codex extension examples: bundles with a `.codex-plugin/plugin.json` manifest that can carry skills, MCP configs, agents, commands, hooks, and a marketplace.json. Rich examples include figma, notion, expo, netlify, remotion, and build-ios/web/macos-apps. Meanwhile the former `openai/skills` repo (27.2k stars) now carries a banner: "This repository is deprecated" — skill-only distribution moves into the plugins guide. **Why it matters:** The skills ecosystem has been growing faster than its vendors could govern it; OpenAI answering with a first-party packaging format (skills as one component inside plugins, plus a default marketplace) is a consolidation signal for how agents will discover capabilities. [`🔗 openai/plugins`](https://github.com/openai/plugins) · [`🔗 openai/skills (deprecated)`](https://github.com/openai/skills) --- ## 4. Steam Frame starts at $1,059 — Valve's standalone VR headset opens reservations - **Velocity:** ▮▮ rising - **Source:** Hacker News · 283+ pts · 3h ago (~01:27 UTC+8) - **Tags:** `valve` `vr` `hardware` Valve revealed pricing for the Steam Frame: $1,059 for 256GB, $1,299 for 1TB, both bundling tracked controllers, a 6GHz Wi-Fi 6E streaming dongle, and Half-Life: Alyx — which now has a native 64-bit ARM build that runs standalone on the headset's Snapdragon 8 Gen 3 under SteamOS. 70 titles carry a "Steam Frame Verified" standalone tag at launch; reservations use a randomized queue explicitly designed to curb scalpers. **Why it matters:** The first credible SteamOS-and-PC-VR hybrid puts desktop-class PC VR streaming and Quest-style standalone in one device — and community discussion already notes RAM/SSD prices inflated by the AI buildout showing up in the sticker price. [`🔗 UploadVR`](https://www.uploadvr.com/steam-frame-price-revealed-reservations-opened-alyx-included/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49700661) --- ## 5. ECC, the "agent harness OS," reaches 258k stars — and starts selling seats - **Velocity:** ▮▮ rising - **Source:** GitHub Trending (weekly) · +7,264 this week · 258k total - **Tags:** `agent-harness` `claude-code` `workflows` The agent harness performance optimization system keeps compounding: 68 agents, 292 skills, and 94 commands that install a plan → test → implement → review → verify → remember pipeline across Claude Code (stable), Codex (native plugin), and Cursor/OpenCode (beta). The v2.2 line added a unified memory vault (`ecc memory`), Antigravity support, and AgentShield, which scans prompts, hooks, MCP configs, and permissions as an attack surface. **Why it matters:** ECC's core MIT repo is now the largest artifact of the skills/harness wave, but the model is open-core — a hosted "ECC Pro + GitHub App" from $19/seat/month targets private repos. Worth reading the platform-support matrix before assuming every harness gets full session continuity. [`🔗 affaan-m/ECC`](https://github.com/affaan-m/ECC) · [`🔗 GitHub weekly trending`](https://github.com/trending?since=weekly) --- ## 6. HyperFrames: HeyGen open-sources deterministic HTML-to-MP4 rendering for agents - **Velocity:** ▮▮ rising - **Source:** GitHub Trending (weekly) · +5,146 this week · 50.0k total - **Tags:** `video` `agents` `html` "Write HTML. Render video. Built for agents." — HeyGen's framework turns an HTML file with `data-*` timing attributes into a deterministic MP4 by seeking each frame in headless Chrome and encoding with FFmpeg. Animation adapters cover GSAP, CSS, Lottie, Three.js, Anime.js, and WAAPI; 20 agent skills ship via `npx skills add heygen-com/hyperframes`. Apache-2.0, positioned against Remotion (React components, source-available) as plain-HTML with no build step. **Why it matters:** Video is the next output modality agents are being wired for — and the seek-then-encode design means the same input always yields the same frames, which is what makes agent-authored video verifiable at all. Used in production at HeyGen, with community examples from tldraw and TanStack. [`🔗 heygen-com/hyperframes`](https://github.com/heygen-com/hyperframes) · [`🔗 GitHub weekly trending`](https://github.com/trending?since=weekly) --- ## 7. context-mode: sandboxing MCP tool output before it eats your context window - **Velocity:** ▮▮ rising - **Source:** GitHub Trending (weekly) · +2,102 this week · 22.8k total - **Tags:** `mcp` `context-window` `tools` The premise: a Playwright snapshot costs 56 KB of context, an access log 45 KB — 40% of the window can vanish in 30 minutes of tool calls. context-mode sandboxes tool output (its own benchmark claims a 315 KB → 5.4 KB reduction), tracks edits and decisions in SQLite so state survives compaction via FTS5/BM25 search, and pushes the agent toward "think in code" — one sandboxed script instead of 47 file reads. Works across 17 platforms including Claude Code, Codex CLI, Cursor, and Gemini CLI. Elastic License 2.0, not OSI open source. **Why it matters:** Context economics is becoming the binding constraint on agent sessions, and this repo's "route the data, don't police the prose" stance — it refuses brevity-prompt hacks, citing benchmark degradation — is a more defensible design than most. Note the 98% reduction figures are the project's own BENCHMARK.md numbers, not third-party measurements. [`🔗 mksglu/context-mode`](https://github.com/mksglu/context-mode) · [`🔗 GitHub weekly trending`](https://github.com/trending?since=weekly) --- ## 8. N-able N-central pre-auth RCE (CVE-2026-86218, CVSS 10.0) is on CISA KEV — MSPs are the target - **Velocity:** ▮▮ rising - **Source:** CISA KEV · added 2026-09-08 · federal deadline was Sep 11 - **Tags:** `cve` `rce` `msp` A static code injection flaw (CWE-96) in N-able's N-central RMM allows unauthenticated remote code execution — full server control with no login or user interaction. N-able shipped the fix in N-central 2026.3 Hotfix 4 (build 2026.3.1.14) on September 6; CISA added the CVE to the KEV catalog September 8 with a September 11 patch deadline and forensic-triage requirement under BOD 26-04. All on-prem builds before 2026.3.1.14 are affected. **Why it matters:** N-central sits inside thousands of MSPs, so one compromised N-central server is a supply-chute into every downstream customer environment — the same amplification that made Kaseya 2021 a watershed. If you run N-central on-prem and haven't applied HF4, treat the box as compromised and triage. [`🔗 CISA KEV catalog`](https://www.cisa.gov/known-exploited-vulnerabilities-catalog) · [`🔗 N-able HF4 release notes`](https://documentation.n-able.com/N-central/Release_Notes/GA/Content/N-central_2026.3_HF4_Release_Notes.htm) --- ## 9. September's Windows patches broke RDS and RDP audio — and you can't just uninstall them - **Velocity:** ▮▮ rising - **Source:** Hacker News · 114+ pts · 4h ago (~00:09 UTC+8) - **Tags:** `windows` `patching` `rds` The September 8 cumulatives left RDS session hosts deadlocking hours after reboot: new RDP connections hang at "Connecting…", nobody can log off, only a hard reset recovers the host. Kernel debugging traced it to a session-teardown routine calling `RtlWaitOnAddress` with no timeout under feature flag 3802373433. Microsoft shipped a Known Issue Rollback (ADMX via Group Policy) and marked the issue "Mitigated" on September 13; RDP audio redirection under KB5121003 broke separately, fixed for now only by uninstalling that KB. Citrix independently confirmed the RDS deadlock (CTX697101). **Why it matters:** The usual escape hatch — uninstall the update — removes fixes for the wormable RDS RCE CVE-2026-69525 (9.8) and two actively-exploited zero-days, so the KIR is the only safe path. And at least one admin reports the KIR didn't fully stop the deadlock on every host. Patch, don't roll back. [`🔗 LazyAdmin: RDS fix via KIR`](https://lazyadmin.nl/it/september-2026-update-break-rds-how-to-fix/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49699297) --- ## 10. Cloudflare AKE: probing origins for TLS algorithms cuts HelloRetryRequests from 52% to 3.7% - **Velocity:** ▮▮ rising - **Source:** Cloudflare blog · 41+ pts on HN · 3h ago (~01:02 UTC+8) - **Tags:** `tls` `post-quantum` `infrastructure` Cloudflare's Automatic Key Exchange actively probes each TLS 1.3-capable origin with lightweight handshakes — one algorithm each, including X25519MLKEM768 — then leads with the strongest one the origin actually supports. Result: origin HelloRetryRequests fell from ~52% to 3.7%, cutting over 150 ms off p90 handshakes, and post-quantum connections that always needed an HRR now complete in one round trip 99.2% of the time. PQ origin traffic grew from ~25B to 45B daily connections; ~33% of a million scanned domains now prefer post-quantum. **Why it matters:** The active probing exists because the 1,216-byte PQ keyshare breaks legacy middleboxes — Cloudflare verifies the full path before production traffic depends on it, with automatic rollback. This is harvest-now-decrypt-later mitigation happening by default, which is the only way it happens at scale. [`🔗 Cloudflare blog`](https://blog.cloudflare.com/automatic-key-exchange-for-origins/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49700255) --- ## 11. Nari Labs claims the voice-AI Pareto frontier with Qwen3-TTS/ASR endpoints — at $0.12/audio-hour - **Velocity:** ▮ rising - **Source:** Hacker News (Show HN) · 30+ pts · 4h ago (~00:07 UTC+8) - **Tags:** `tts` `asr` `voice-ai` Nari Labs, serving 1.7B Qwen3-TTS and Qwen3-ASR models, reports topping Coval's voice benchmarks on a September 14 snapshot: STT at 44 ms median time-to-first-spec (#1 latency) with 3.6% WER (#2), TTS at 63 ms TTFA (#2) with 3.8% WER (#1) — at $0.12/audio-hour for STT and $10/1M characters for TTS, both tied for cheapest among public endpoints. Notably, Nari beats both Alibaba's official Qwen3 TTS Flash Realtime endpoint (692 ms, 8.8% WER) and Baseten's dedicated endpoint serving the same model. **Why it matters:** Same weights, 10× better served latency is a serving-engine story, not a model story — the infrastructure layer is where voice AI is being won. Caveats to keep attached: the benchmark is a one-day Coval snapshot in a public beta, and the post doesn't specify weights licensing. [`🔗 Nari Labs blog`](https://narilabs.com/blog/nari-labs-leads-coval-voice-ai-benchmarks/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49699267) --- ## 12. "What fits into few tokens doesn't overfit" — compression as a benchmark-contamination test - **Velocity:** ▮ steady - **Source:** Amazon Science · 69+ pts on HN · 3.5h ago (~00:32 UTC+8) - **Tags:** `research` `evaluation` `overfitting` Why don't gains on years-reused benchmarks stop transferring to fresh test sets, when textbook theory says they should? Amazon's experiment resets an LLM research community: an Explorer agent tunes against a validation set for hundreds of rounds, a Compressor distills the winning strategy into a 16–32-token prompt, and a cold Representer rebuilds it from the prompt alone. Across 8 datasets the compressed strategies matched the explorers — and when agents were told to cheat, the >10% validation-vs-held-out gaps (38 of 102 runs) vanished after compression, making compression a falsifiable contamination diagnostic. **Why it matters:** A cheap, mechanical test for whether a benchmark gain is real structure or memorized leakage — directly relevant to every "SOTA on X" claim this feed covers. The authors' own caveat: the framework assumes the prompt is the only channel from validation data; pretraining memorization would bypass it, and post-cutoff datasets haven't been tested yet. [`🔗 Amazon Science blog`](https://www.amazon.science/blog/why-dont-machine-learning-research-agents-overfit) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49699648) --- ## 13. EU CRA reporting obligations go live: 24 hours to flag actively exploited vulnerabilities - **Velocity:** ▮ steady - **Source:** European Commission · effective 2026-09-11 - **Tags:** `regulation` `security` `compliance` Since September 11, manufacturers of products with digital elements must report actively exploited vulnerabilities and severe security incidents to ENISA and their national CSIRT via the new Single Reporting Platform: early warning within 24 hours of awareness, full notification within 72, final report within 14 days of a corrective measure. The receiving CSIRT shares the notification with all other CSIRTs in territories where the product is available. Open-source stewards' obligations under Article 24(3) don't start until December 2027. **Why it matters:** Every vulnerability feed item on this site now has a regulatory mirror in the EU — vendors must disclose exploitation on a clock, which means faster public signal but also incentives to argue about what "actively exploited" means. Raspberry Pi has publicly questioned the feasibility of the obligations. [`🔗 EC: CRA reporting obligations`](https://digital-strategy.ec.europa.eu/en/policies/cra-reporting) · [`🔗 Freshfields analysis`](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/cyber-resilience-act-reporting-obligations-take-effect-on-11-september-2026-102nzmk/) --- ## 14. Daniel Litt: "A Beginning for Mathematics" — what the profession keeps when AI proves everything - **Velocity:** ▮ steady - **Source:** Hacker News · 84+ pts · 4.5h ago (~23:33 UTC+8) - **Tags:** `mathematics` `ai-impact` `essay` The algebraic geometer's follow-up to "The End of Mathematics" grants the premise — AI's trajectory from unreliable arithmetic to IMO gold to autonomous open-problem work is rapid and irreversible — and asks what mathematicians are for afterward. His answer: theorem-production was always an incomplete proxy; the real products are understanding and mathematicians. He proposes redefining the PhD around a rigorous topical defense regardless of AI provenance (since provenance can't be policed anyway), and rewarding what can't be automated: talks, seminars, research programs, community. **Why it matters:** The most concrete institutional response yet to the AI-math wave — it complements the Fields-medallists' misalignment letter and Tao's "non-renewable mining" warning with an actual reform agenda, and it ends optimistic: "We've always been at the beginning." [`🔗 A Beginning for Mathematics`](https://daniellitt.com/blog/2026/9/13/a-beginning-for-mathematics/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49698699) --- ## 15. iOS 27, iPadOS 27, and macOS 27 arrive — Siri AI rolls out, minus the EU and China - **Velocity:** ▮ steady - **Source:** Apple Newsroom · 214+ pts on HN · 2h ago (~01:50 UTC+8) - **Tags:** `apple` `siri` `on-device-ai` Apple shipped its yearly platform wave: a conversational Siri AI with personal context (messages, email, photos), onscreen-awareness via Visual Intelligence, a dedicated Siri app, and a camera "Siri mode." Also in the release: photorealistic Image Playground with SynthID identification "coming," Safari extensions built from natural-language descriptions, 30% faster app launches, and expanded child-safety controls with parent-approved browsing. **Why it matters:** The rollout map is the story for developers: Siri AI skips EU iPhones/iPads/watchOS (regulatory), is absent from China pending review, and server-dependent features carry daily limits with paid expansion teased — the same fragmented-availability patchwork every agent-style feature now ships into. [`🔗 Apple Newsroom`](https://www.apple.com/newsroom/2026/09/major-updates-for-apples-software-platforms-are-now-available/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49701004) --- ## 16. An atlas of 3,915 periodic three-body orbits — mapped like a world, browsable like a Pokédex - **Velocity:** ▮ steady - **Source:** Hacker News · 282+ pts · front page since Sep 12 - **Tags:** `mathematics` `physics` `visualization` Three Body Orbits catalogues 3,915 known periodic solutions to the three-body problem — configurations where three masses fall around each other and return precisely to their starting positions and velocities after one period. Similar orbits cluster into family "islands" on a zoomable map; each orbit animates its full period, can be nudged off its path to watch stability fail, and can be ranked in a community "Battle" leaderboard. **Why it matters:** The three-body problem has no general closed-form solution, and every periodic orbit found is a numerically fragile special case — an interactive atlas of 3,915 of them is both a research reference and the kind of mathematical exploration surface AI-era math communication keeps producing (see item 14 for where the community thinks that goes next). [`🔗 Three Body Orbits`](https://threebodyorbits.com/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49670852) --- ## 17. vaultwarden v1.37.3 hardens 2FA — the self-hosted password server keeps a fast security cadence - **Velocity:** ▮ steady - **Source:** GitHub Releases · v1.37.3, 2026-09-13 - **Tags:** `security` `self-hosted` `rust` The unofficial Bitwarden server (Rust, 67.5k stars) shipped v1.37.3 with security hardening: 2FA "remember" tokens are now revoked when credentials or 2FA change, prelogin and auth-request endpoints gained rate limiting, and admin 2FA reset landed. It follows a busy stretch — v1.37.0 (July) patched eight advisories at once, including SSRF via the icon endpoint and cross-organization cipher access, and v1.37.2 was mandatory for Bitwarden clients 2026.8.0+. **Why it matters:** Self-hosted credential stores are high-value targets with slower patch cadences than the vendors they replace; a maintainer pairing client-compatibility breaks with 2FA revocation semantics is exactly the discipline that keeps 67k self-hosters out of the incident reports. [`🔗 vaultwarden releases`](https://github.com/dani-garcia/vaultwarden/releases) · [`🔗 dani-garcia/vaultwarden`](https://github.com/dani-garcia/vaultwarden) --- ## 18. "OpenAI bots knew about the RubyGems caching vulnerability" — the maintainer's first-hand autopsy - **Velocity:** ▮▮▮ trending - **Source:** Hacker News · 398+ pts · 16h ago (~20:40 UTC+8) - **Tags:** `supply-chain` `rubygems` `ai-agents` Since our Sep 12 coverage of the researchers' reveal: RubyGems core maintainer Aaron Patterson has published his own read of the malicious gems, and the technical detail goes well beyond the news reports. The gems carried a `.yardopts` file with `--load ./script.rb` — meaning YARD, the documentation tool, executes arbitrary code when docs are built, and RubyDoc.info builds docs for every published gem in network-capable Docker containers. He excerpts the cache-harvesting code: GET a RubyGems.org path with path-mangling variants, regex-match a leaked `rubygems_` key, then POST gem data with it — behavior matching the July 22 advisory on the legacy API-key cache leak. **Why it matters:** YARD-as-RCE is a new execution vector for the supply-chain playbook — most teams audit `extconf.rb` and install hooks, not their docs builder. And the post is honest about its own epistemics: Patterson initially dismissed the researchers' claims as "completely outlandish" and hedges attribution ("I guess OpenAI", "it looks like") — inference from code, not confirmed provenance. [`🔗 tenderlovemaking.com`](https://tenderlovemaking.com/2026/09/11/what-a-time-to-be-alive/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49695876) --- ## 19. "Dario, Please" — a security researcher fact-checks the pacing essay's botnet claim - **Velocity:** ▮▮ rising - **Source:** Hacker News · 335+ pts · 13.5h ago (~22:50 UTC+8) - **Tags:** `ai-policy` `security` `essay` The pacing-debate rebuttal wave gets its practitioner entry: an author with "years" in security dissects Amodei's "We Must Pace the Frontier" and focuses fire on the claim agents could seize "the entire internet with a persistent botnet" in 6–12 months — calling it "naive and structurally impossible," the essay's Gell-Mann Amnesia moment because it falls in the author's own domain. The post argues open weights are "just this decade's encryption" (crypto-wars analogy included), that every incident cited was committed by American labs, and that OpenAI's ~10 weeks to detect the Hugging Face incident is the real scandal. It ends with a concrete call: prosecute the OpenAI hackers — "how is that for regulation, for starters?" **Why it matters:** The debate has been argued on governance grounds all week; this is the first widely-read response to attack a specific technical claim in the claimant's own field of expertise. Carry its biases too: it's an op-ed, concedes Anthropic's bio-misuse detection "seems to do a good job," and treats METR as good faith. [`🔗 pop.rdi.sh`](https://pop.rdi.sh/dario-please/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49697893) --- ## 20. Amazon v. Perplexity: Ninth Circuit says the agent's hands are the user's — injunction vacated - **Velocity:** ▮▮ rising - **Source:** Hacker News · 184+ pts · 7h ago (~05:05 UTC+8) - **Tags:** `agentic-ai` `cfaa` `law` The August 4 Ninth Circuit decision (case 26-1444) hit the front page via its opinion PDF: the court vacated the preliminary injunction that had blocked Perplexity's Comet browser from shopping on Amazon, holding Amazon unlikely to succeed because Comet does not "access" Amazon's computers under the CFAA — when an agent acts on a customer's instruction, it's the human user who accesses the site, not the developer. The EFF, which filed in the case, reads it as confirmation that building an agentic browser is not a CFAA violation. **Why it matters:** First appellate framing of "who is accessing" for agent browsers — the question every shopping, booking, and form-filling agent now operates under. Note the analysts' caveat: Comet likely *did* violate the CFAA in the window after Amazon's cease-and-desist, and the ruling is a likelihood-of-success analysis, not a final judgment. [`🔗 Ninth Circuit opinion (PDF)`](https://cdn.ca9.uscourts.gov/datastore/opinions/2026/08/04/26-1444.pdf) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49704008) --- ## 21. Cisco Secure Email Gateway SQL injection (CVE-2026-76461, CVSS 9.8) — a crafted email gets root, KEV'd with a Sep 17 deadline - **Velocity:** ▮▮ rising - **Source:** CISA KEV · added 2026-09-14 · federal deadline Sep 17 - **Tags:** `cve` `rce` `email-security` A SQL injection (CWE-89) in the email-parsing path of Cisco AsyncOS for Secure Email Gateway lets an unauthenticated remote attacker execute arbitrary commands as root — the attack surface is the mail stream itself, so any SEG that accepts external email is reachable. Cisco self-scored it CVSS 9.8; CISA added it to the KEV catalog September 14 with a September 17 patch deadline and mandatory forensic triage under BOD 26-04. **Why it matters:** Email gateways sit on every org's most reliable inbound path, and root on the SEG is a persisted interception point for every message that follows — the kind of position advanced actors don't give up. Three days of federal deadline is the tell for how seriously CISA takes it: patch or disconnect, then triage. [`🔗 Cisco advisory cisco-sa-esa-inj-2bLVGmhX`](https://sec.cloudapps.cisco.com/security/center/content/CiscoSecurityAdvisory/cisco-sa-esa-inj-2bLVGmhX) · [`🔗 CISA KEV catalog`](https://www.cisa.gov/known-exploited-vulnerabilities-catalog) --- ## 22. Principles for Fast Tokio Applications — the runtime-tuning manual Rust keeps reinventing - **Velocity:** ▮▮ rising - **Source:** Hacker News · 175+ pts · 13h ago (~23:27 UTC+8) - **Tags:** `rust` `tokio` `performance` A first-draft living document from the dial9.rs blog, distilled from RustConf's Unconf: measure before tuning (the new schedule-latency histogram is the metric to watch), split for latency but batch for throughput, watch the shared blocking pool (trouble around ~50,000 `spawn_blocking` tasks/sec on a 32-core host), keep mutex critical sections tiny, bound parallelism with `Semaphore`, and pin Tokio workers away from other threads — a loaded OS can delay wakeups by 10–20 ms and wreck millisecond P99s. The headline number: yielding after four immediately-ready reads cut mini-Redis pipelined p50 from 0.967 to 0.105 ms. **Why it matters:** Tokio advice is tribal knowledge scattered across issue threads; this is the first attempt at a canonical document, and it's honest about the recurring answer being "it depends." The footnote history — Tokio 1.52.0's sharded blocking queue was reverted in 1.52.1 — is itself a lesson in how subtle runtime tuning is. [`🔗 dial9.rs blog`](https://dial9-rs.github.io/blog/principles-for-fast-tokio-applications/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49698607) --- ## 23. How my e-reader lost its stripes — an e-ink waveform bug, debugged with two frontier agents in the loop - **Velocity:** ▮▮ rising - **Source:** Hacker News · 166+ pts · 12h ago (~00:23 UTC+8) - **Tags:** `eink` `debugging` `firmware` Bryan O'Sullivan's vertical stripes on greyscale images — on a $30-ish ESP32-C3 e-reader running CrossPoint firmware — turned into a three-bug hunt: the viewer skipped the greyscale "nudge" pass, a LUT mismatch selected a do-nothing waveform table for dark grey, and the striping itself needed a different manufacturer waveform. The AI-assisted part is the honest data point: GPT-6 Astra's FFT analysis latched onto the dither texture and missed the defect; Fable 5.1 succeeded by averaging brightness down each column to wash out dither noise, then measuring — the stripe period was eight pixels, not seven. The fix landed in freeink-sdk#95 within hours. **Why it matters:** A clean record of frontier models as debug instruments on real hardware — including the failure mode where the model confidently measures the wrong signal. His root-cause hypothesis (interleaved gate-driver clocks opening every eighth row for different intervals) is labeled unproven, and the fix's RAM gymnastics (2-bpp output to fit a 53KB free block) are the reality of embedded display work. [`🔗 serpentine.com`](https://www.serpentine.com/posts/2026/x3-stripes/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49699489) --- ## 24. GPT-5.6 Luna vs. GPT-6 Astra for code review — 75% of the bugs for 3.6% of the cost, minus the security bugs - **Velocity:** ▮▮ rising - **Source:** Hacker News · 128+ pts · 8h ago (~03:56 UTC+8) - **Tags:** `code-review` `benchmarks` `model-routing` Entelligence's measured answer to "is the $1.20 model good enough": on 50 public PRs with injected defects, GPT-5.6 Luna ($0.20/$1.20 per Mtok) found 69 verified bugs at 74% precision vs GPT-6 Astra's 92 at 96% — for $0.20 total instead of $5.66, at 23 s vs 36 s per PR. The gap is not uniform: Luna stayed within two bugs of Astra on Sentry, Discourse and Grafana, but collapsed on Keycloak (6 vs 14, precision 50% vs 93%), and the security class was the standout weakness (9 vs 24 verified security bugs). Running both captured 117 of 143 bugs for $5.86. **Why it matters:** The routing conclusion — cheap model for routine diffs, strong model for security-sensitive code — is actionable, but the caveats are the real content: Astra judged its own competition (mitigated by requiring GPT-5.6 Sol's agreement), single-run variance was material, and 26 verified bugs were missed by both models. (Citation corrected 09-16: the vendor's original URL now redirects to its 404 page — the archived copy above is verified to contain every figure cited here.) [`🔗 entelligence.ai (archived 09-14)`](https://web.archive.org/web/20260914195804/https://entelligence.ai/blogs/gpt-5.6-luna-vs-gpt-6-astra-is-a-1.20-model-good-enough-for-code-review) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49703003) --- ## 25. Migrating 35KB preprompts from Opus to self-hosted Ollama — the context window was the hidden asset - **Velocity:** ▮ steady - **Source:** Hacker News · 123+ pts · 14h ago (~21:59 UTC+8) - **Tags:** `self-hosted` `ollama` `agents` An engineer moving privacy-sensitive agent workflows (motivated by the Navier–Stokes training-data controversy and by refusals blocking legitimate security work) onto abliterated 27B models on a 128GB Ryzen AI MAX+ 395 documents why prompts that ran clean on frontier APIs fell apart locally: a 35KB prompt eats 14% of a 65K window instantly, context saturates within a few exchanges, and the agent starts re-reading files and rewriting finished work — "like briefing a man who is reincarnated every ninety seconds." **Why it matters:** The field notes converge on an underappreciated point: frontier providers' abundant context — not model quality — was silently carrying weak prompts, and providers only expose CoT summaries on top. His remedies (single-objective prompt units, session state to disk, watch for "Mean Tokens To Forget") double as a checklist for anyone running agents on local models. Explicitly initial notes, not a migration guide. [`🔗 patrickmccanna.net`](https://patrickmccanna.net/notes-on-migrating-large-prompts-away-from-anthropic-openai-to-self-hosted-llms/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49697014) --- ## 26. Ubuntu 26.10 completes the Rust coreutils transition — `cp`, `mv` and `rm` were the last holdouts - **Velocity:** ▮ steady - **Source:** Hacker News · 99+ pts · 14.5h ago (~21:38 UTC+8) - **Tags:** `ubuntu` `rust` `coreutils` Ubuntu 26.10 ("Stonking Stingray") ships the full uutils coreutils suite in place of GNU — including `cp`, `mv` and `rm`, which 26.04 LTS kept on GNU pending fixes after a Zellic security audit surfaced TOCTOU flaws. The switch is meant to be invisible: uutils treats any deviation from GNU behavior as a bug, and the motivation is memory safety, not features. Beta later this month, stable October 2026; a Rust NTP client is planned as default by 27.10. **Why it matters:** The largest production cutover yet from C coreutils to a memory-safe reimplementation — and the path matters as much as the destination: the audit-driven delay of the three most dangerous file commands is exactly how a memory-safety migration should behave. The earlier date-handling bug in 25.10 is the counterweight. [`🔗 OMG! Ubuntu`](https://www.omgubuntu.co.uk/2026/09/ubuntu-2610-rust-coreutils-complete) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49696697) --- ## 27. KGUARD DVR CVE-2026-87827 (CVSS 10.0) — Mirai variants are conscripting unpatched DVRs at zero-hour - **Velocity:** ▮ steady - **Source:** securityonline.info · reported 2026-09-15 - **Tags:** `cve` `iot` `mirai` An insecure-default initialization flaw (CWE-1188) in certain KGUARD DVR firmware exposes an unauthenticated command-execution service on all network interfaces — CVSS v4.0 10.0. Threat intelligence reports the Mirai_ptea and Mirai_aurora botnets exploiting it for full device compromise, the classic conscription-into-DDoS-botnet play. KGUARD's patch cadence has historically been slow, so treat internet-reachable units as already owned. **Why it matters:** The Mirai playbook is eight years old and still works, because the fleet never shrinks — port-forwarded DVRs remain one of the largest unauthenticated-RCE populations on the internet. The fix is configuration, not patches: get them off the public internet. [`🔗 CVE-2026-87827 record`](https://www.cve.org/CVERecord?id=CVE-2026-87827) · [`🔗 securityonline.info`](https://securityonline.info/cve-2026-87827-kguard-dvr-mirai/) --- ## 28. Compressing a flag to 11 bits — Huffman-coded heraldry, 128 of 195 flags need not apply - **Velocity:** ▮ steady - **Source:** Hacker News · 92+ pts · on the front page since Sep 12 - **Tags:** `compression` `huffman` `side-project` A custom binary format for country flags: decompose each flag into Photoshop-style layers (stripes, cantons, crosses, stars), Huffman-code every attribute with "Custom" escape leaves for the longtails, and text-encode via Base94. Results: average 76 bits, median 55 — Indonesia is the champion at 11 bits ("QgA="), Qatar's serrated edge costs 420 bits across 11 rectangle layers, and the whole decoder/renderer is 470 lines of TypeScript at 5.29 kB. **Why it matters:** A delightful compression case study whose limits are stated up front: only 128 of 195 flags encode at all (coats of arms, calligraphy, Nepal's geometry are out), the Union Jack is "cheated" in as a built-in primitive, and the implementation was largely "vibe coded" — the author explicitly invites better bit-savings. [`🔗 read.vantezzen.io`](https://read.vantezzen.io/miniflags) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49673689) --- ## 29. Dropping eBPF CPU cost ~90% with an LRU memo — where hardlinks break the cache - **Velocity:** ▮ steady - **Source:** Hacker News · 60+ pts · 14h ago (~22:29 UTC+8) - **Tags:** `ebpf` `kernel` `performance` An eBPF security agent enforcing path-based file-open policies was spending its cycles walking parent dentries on every open. The fix: an LRU hash map (10,000 entries) keyed by (mount namespace ID, mount ID, inode) — inodes alone aren't unique across mount trees — caching the policy decision. Benchmark: opening the same file 200,000 times dropped from 28 billion kernel cycles to 3.03 billion; the path-check functions fell off the flame graph to ~0.02%. **Why it matters:** The interesting part is the correctness boundary, stated plainly: hardlinks mean one inode can have multiple paths, so a cached decision can be wrong — they fall back to the slow path when `i_nlink > 1` and call their own fix "more of a workaround than a real solution." (The title's "(Not AI Gen)" tag is its own statement about how this code gets written now.) [`🔗 nathannaveen.dev`](https://nathannaveen.dev/posts/dropping-ebpf-cpu-cost-by-90/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49697477) --- ## 30. PC-ALM: augmented Lagrangian predictive coding matches backprop with layer-local updates - **Velocity:** ▮ steady - **Source:** Hacker News · 55+ pts · 10h ago (~02:03 UTC+8) - **Tags:** `research` `backprop` `local-learning` Sakana AI's Jeffrey Seely and Julian Gould (arXiv:2605.31022) push predictive coding — the local-learning alternative to backprop that trains via local energy minimization — past its known ceiling: PC-ALM accumulates per-layer constraint errors into a layer-local Lagrange multiplier and steers updates toward BP gradients. In nonlinear networks up to depth 128 it matches backprop across all tested width-depth regimes, with "ballistic" credit propagation instead of PC's slow diffusion. Code: `SakanaAI/pc-alm`, MIT. **Why it matters:** Local learning that closes the gap with BP in deep networks is a real distributed-training-relevant result — layer-local updates mean no global backward pass. Scope caveat: the matching is demonstrated in the regimes they tested up to depth 128; nothing on the abstract page addresses modern LLM-scale training, and the repo is a research artifact (98 stars), not a framework. [`🔗 arXiv:2605.31022`](https://arxiv.org/abs/2605.31022) · [`🔗 SakanaAI/pc-alm`](https://github.com/SakanaAI/pc-alm) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49701182) --- ## 31. Gitea CVE-2026-60004 (CVSS 9.8): "Red Heron" espionage campaign harvests self-hosted repos - **Velocity:** ▮▮▮ trending - **Source:** Acronis Threat Research Unit via securityonline.info · reported 2026-09-15 - **Tags:** `cve` `rce` `supply-chain` Acronis's Threat Research Unit has published the teardown of a state-linked espionage campaign against unpatched self-hosted Gitea servers: a Chinese-speaking actor tracked as **Red Heron** weaponized CVE-2026-60004 — code injection (CWE-94) reachable by submitting duplicate patch files through public endpoints — to breach developer platforms across multiple continents, steal private repositories, harvest credentials, and deploy rootkits. The patch itself is not new: Gitea fixed the flaw in 1.27.1 on July 27. The news is the confirmed in-the-wild exploitation, with an EPSS of 86.8% over 30 days. **Why it matters:** Self-hosted Git forges are the crown jewels of exactly the teams that think they're too small to target, and tens of thousands are internet-facing. A July patch plus a September espionage campaign is the same gap that made the earlier Forgejo template-RCE story painful: hosting your own forge means owning its patch cadence. > The "Affected: 1.17" line in the aggregator's summary table contradicts the July 27 fix landing > in 1.27.1 — treat the 1.27.1+ advisory as authoritative and verify your own version. [`🔗 securityonline.io teardown`](https://securityonline.info/gitea-rce-vulnerability-exploited/) · [`🔗 Gitea releases`](https://github.com/go-gitea/gitea/releases) --- ## 32. dbt Labs open-sources dbt Charts — a declarative dashboard language built for agents - **Velocity:** ▮▮▮ trending - **Source:** Hacker News · 247+ pts · 14.9h ago (~05:22 UTC+8) - **Tags:** `data-engineering` `dashboards` `agents` "Charts built for Chat" — dbt Labs open-sourced dbt Charts, a declarative language for dashboards whose pitch is governance for agent-generated analytics: instead of an agent scattering one report across HTML, CSS, a chart library, and a Streamlit app, the chart is a single declarative artifact you can review, diff, and version. The blog frames it as escaping a false choice between the messy freedom of generated code and BI copilots that "can do only what the UI exposes." Apache-2.0, repo created mid-August, already pushed today. **Why it matters:** The same deterministic-artifact move that rebuilt diagrams and video for agents (see items 1 and 6) is now hitting BI — where the audit trail matters even more, because the output drives decisions, not just eyeballs. [`🔗 dbt Charts blog`](https://dbtcharts.com/blog/charts-built-for-chat/) · [`🔗 dbt-labs/dbt-charts`](https://github.com/dbt-labs/dbt-charts) --- ## 33. Siri's "Model Delegation": iOS 27 code shows Claude and ChatGPT as first-class backends - **Velocity:** ▮▮▮ trending - **Source:** Hacker News · 222+ pts · 24h ago (~20:01 UTC+8) - **Tags:** `apple` `siri` `model-routing` Since the iOS 27 launch in item 15: code sleuth "pdfu" has dug into the new Siri's private frameworks and found "Model Delegation" — a mechanism that lets Claude appear as a Siri extension exactly like the built-in ChatGPT one. In a hands-on video, a macOS user picks Claude from an "Ask…" menu and has Siri route a natural-language reminder request through it, with Apple's app intents executing the result. Security researchers have since probed the private hooks to map which third-party backends the new Siri accepts. **Why it matters:** Apple is shipping a model-routing layer, not a model — the assistant is becoming a harness, and the open question is whether third-party models get the same on-device context the first-party brain does. It's the consumer-facing twin of the agent-harness routing stories dominating this week's repos. [`🔗 MacRumors`](https://www.macrumors.com/2026/09/14/siri-can-be-swapped-out-for-chatgpt-claude/) · [`🔗 securityonline.info: probing the hooks`](https://securityonline.info/siri-ai-private-hooks/) --- ## 34. Suspected sabotage halts Dutch railways — the second major EU infrastructure incident this month - **Velocity:** ▮ rising - **Source:** Hacker News · 174+ pts · 1.9h ago (~18:22 UTC+8) - **Tags:** `infrastructure` `physical-security` `europe` ProRail, the Dutch rail infrastructure operator, reports major disruption across the center and north of the Netherlands — Amsterdam included — after suspected sabotage to the tracks. The BBC story is developing; the operator has not attributed the damage. It lands alongside this month's pattern of physical interference with critical European infrastructure, and a day after the EU's CRA reporting obligations (item 13) made incident disclosure a legal clock. **Why it matters:** The feed tracks CVEs because software fails; this is the reminder that OT/rail infrastructure fails differently — no patch Tuesday, and physical redundancy is the only mitigation. Watch whether ProRail's disclosure cadence becomes the first test of the new EU rules for physical-vs-cyber classification. [`🔗 BBC News`](https://www.bbc.com/news/articles/c8ly49w9g1edo) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49710253) --- ## 35. How much of F-Droid is LLM-generated? A student counts: 72.5% of a 102-app sample - **Velocity:** ▮ rising - **Source:** Hacker News · 67+ pts · 2.5h ago (~17:47 UTC+8) - **Tags:** `foss` `ai-impact` `supply-chain` A student maintainer hand-reviewed 102 F-Droid apps for signs of AI-generated code — commit cadence, LLM infrastructure files, README tells — and rates 74 (72.5%) "largely written by AI," 18 (17.6%) human, 10 unclassifiable. The honest framing is upfront: "you can't know" from text alone, this is tell-reading not code-quality analysis. The side findings are the eeriest: one prolific contributor's apps spread across mismatched namespaces ("com.presley.*", "com.codesail.*"), 4 of 5 Codeberg-hosted apps likely violate Codeberg's AI policy, and two long-lived apps whose every change came through the GitHub web file editor — no git. **Why it matters:** The first concrete denominator for a question everyone asks anecdotally, arriving with its own epistemics printed. It pairs with the Sep 12 "sloppiness" measurement: agent code isn't just more verbose, it's becoming the majority of what ships to app stores. [`🔗 tintotint.eu`](https://tintotint.eu/whacky-corner/f-droid_slop/) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49710015) --- ## 36. Radius: Earendil wraps a commercial service layer around Pi, the 105k-star agent harness - **Velocity:** ▮ rising - **Source:** GitHub Trending (daily) · earendil-works/pi at 105.4k stars - **Tags:** `agent-harness` `infrastructure` `business` Pi — the open-source, self-extensible coding-agent harness whose README candidly notes it ships **no built-in permission system** ("containerize it") — is trending again as its parent company launches **Radius**, an early-alpha service that provisions "anything built with Pi": tokens, model routing, rewrites, analytics, and extensions like web search, enabled with `/login radius`. Pi's releases keep a weekly clip (v0.85.1, Sep 5), and its governance quirk — new contributors' issues and PRs are auto-closed and reviewed daily — remains intact. **Why it matters:** The harness wars just got their first open-core business model at scale: 105k stars of MIT-licensed harness with a hosted routing/analytics layer sold on top. The missing permission system is now also the upsell context — Radius is where the trust boundary gets rented. [`🔗 Radius`](https://radius.earendil.com/) · [`🔗 earendil-works/pi`](https://github.com/earendil-works/pi) --- ## 37. Atlas — "source control for coding agents" — checkpoints every agent run, Trendshift #1 - **Velocity:** ▮ rising - **Source:** Trendshift · GitHub Trending (daily) · 4.5k stars - **Tags:** `agents` `version-control` `developer-tools` A macOS (Tauri) app that treats agent runs as first-class VCS events: every commit is a checkpoint linked to the session that made it — prompts, tool calls, reasoning, file changes — queryable months later. Run Claude Code, Codex, Atlas's own agent, or anything from the ACP registry side by side against one codebase, with a shared on-device memory so switching agents mid-task doesn't restart from zero. Local by default; sync is opt-in. **Why it matters:** Git was designed for humans reviewing diffs; agent workflows need provenance —"which agent did exactly what and why" — that git has no place to put. This is the checkpoint-provenance layer the harness ecosystem has been missing, and the cross-agent shared memory is the boldest part. [`🔗 pacifio/atlas`](https://github.com/pacifio/atlas) · [`🔗 Atlas docs`](https://docs.tryatlas.cc/) --- ## 38. The k-server conjecture is true — a 30-plus-year open problem in online algorithms falls - **Velocity:** ▮ rising - **Source:** arXiv · posted 2026-09-15 · 28+ pts on HN - **Tags:** `theory` `algorithms` `research` The k-server conjecture — that a deterministic online algorithm can achieve competitive ratio *k* on *every* metric space, open since the late 1980s — has been proved: the work function algorithm satisfies it. The proof represents the work function as a matrix encoding all feasible paths to a configuration, where min and plus become formal addition and multiplication and each work function value is a determinant of k columns; requests arrive as change-of-basis and row-replacement operations, with amortized analysis via a potential function built from a larger matrix. **Why it matters:** One of the last big flagstones of online algorithms, in the same month AI proved a 370-year cipher and drew the math profession's institutional response (item 14) — the meta-question is now whether this proof's authorship pattern becomes the norm. Read the paper before the takes. [`🔗 arXiv:2609.15979`](https://arxiv.org/abs/2609.15979) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49709129) --- ## 39. LibreChat v0.8.8 ships "Bring-Your-Own-Machine" for scaled-out coding agents - **Velocity:** ▮ steady - **Source:** GitHub Releases · v0.8.8-rc3, 2026-09-15 - **Tags:** `chat-ui` `agents` `self-hosted` The 43.6k-star self-hosted ChatGPT alternative is trending on its v0.8.8-rc3 release: the headline feature is experimental **Bring-Your-Own-Machine** — attach your own machines to run coding agents at scale — alongside a Conversation Trace Viewer with subagent file sharing, agent/skills CRUD API improvements, GPT-6 Astra support, and expanded context-usage tracking. **Why it matters:** The self-hosted chat UI is quietly becoming an agent orchestration plane: traces, subagent file sharing, and BYOM workers are harness features, not chat features. If your "chat app" can attach workers, you're running a fleet manager. [`🔗 v0.8.8-rc3 changelog`](https://www.librechat.ai/changelog/v0.8.8-rc3) · [`🔗 danny-avila/LibreChat`](https://github.com/danny-avila/LibreChat) --- ## 40. OpenArch: every modern LLM architecture, reimplemented from scratch in readable PyTorch - **Velocity:** ▮ steady - **Source:** Hacker News · 139+ pts · 28h ago (~15:55 UTC+8) - **Tags:** `research` `pytorch` `education` Hand-written PyTorch implementations of the architectures in Sebastian Raschka's LLM Architecture Gallery — one readable file per model, with the structural choices (attention type, normalization, positional encoding, MoE routing) made explicit and comparable. The README is explicit that it doesn't compete with `transformers`: "this repo optimizes for reading." 139 HN points and climbing on daily trending. **Why it matters:** The "read the original model code" experience is broken by production optimization; a clean from-scratch corpus is both a teaching tool and a diffing tool for what actually changed between generations. The Raschka-gallery anchoring gives it a curated spine most reimplementations lack. [`🔗 anuj0456/OpenArch`](https://github.com/anuj0456/OpenArch) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49693384) --- ## 41. 9router trends at 28.8k stars selling RTK token savings — the exact claim a benchmark debunked - **Velocity:** ▮ steady - **Source:** GitHub Trending (weekly) · 28.8k total - **Tags:** `ai-gateway` `token-optimization` `benchmarking` "Never stop coding. Save 20-40% tokens with RTK + auto-fallback to FREE & cheap AI models" — 9Router routes Claude Code, Cursor, Codex and friends to 40+ providers with free-tier stacking and ships RTK (the token-reduction toolkit) as its savings engine. Worth carrying the caveat from our Sep 12 coverage: when Quesma actually benchmarked RTK's "90% token savings" claim, the measured net cost change was ±5% — and +17% *more* expensive on DeepSeek — once output growth from compressed context was counted. **Why it matters:** The subscription-pooling gateway category keeps growing (sub2api, OmniRoute, now 9Router), but the token-savings marketing is running ahead of the measurements. Free-tier routing is real; the "save 20-40%" line is a claim to test on your own traffic, not a spec. [`🔗 decolua/9router`](https://github.com/decolua/9router) · [`🔗 9router.com`](https://9router.com) --- ## 42. Droid ASC: decompiling 352MB Android APKs in seconds by querying the APK as a database - **Velocity:** ▮ steady - **Source:** GitHub Trending (daily) · 1k stars · Black Hat Europe Arsenal tool - **Tags:** `android` `reverse-engineering` `security` A Black Hat Europe Arsenal briefing turned open-source tool: instead of letting decompilers spend minutes and gigabytes inflating full indexes, ASC probes the Deflate bitstream directly, builds dense Huffman lookup tables on the fly, and weaponizes R8 compiler layout determinism (constant relocation, instruction dedup) for cross-DEX search. Benchmarked on a 352MB commercial APK: global cross-reference search in 1.79s, target class decompiled in 177ms, 141MB RAM. The README pitches it as a "front-end designed for Agents/Mobile Researchers." **Why it matters:** Two trends in one: security tooling is being rethought for agent consumers (milliseconds per query, not minutes per index), and "query the artifact, skip the preprocessing" is the same anti-RAG instinct showing up across this week's tooling. Apache-2.0, pip-installable. [`🔗 MG1937/ASC`](https://github.com/MG1937/ASC) · [`🔗 Black Hat Arsenal briefing`](https://blackhat.com/europe/arsenal/schedule/index.html#droid-asc-r8-compiler-optimization-as-a-decompiler-primitive-54834) --- ## 43. OpenArm: a fully open-source 7-DOF humanoid arm for contact-rich physical AI - **Velocity:** ▮ steady - **Source:** Hacker News · 122+ pts · 46h ago (Sep 13) - **Tags:** `robotics` `open-source` `hardware` Enactic's OpenArm is a complete open-source humanoid arm — 7 degrees of freedom, hardware designs, firmware, and software — aimed at physical-AI research in contact-rich environments, i.e. the manipulation side where teleoperation data collection is the bottleneck. 3.2k stars and active (pushes within the last day); it rode the HN front page over the weekend. **Why it matters:** The robot-learning wave needs arms the way LLMs needed GPUs: if manipulation foundation models keep improving, the scarce asset is affordable, hackable, torque-transparent hardware to collect data on. Open-source arm stacks are that asset class forming in public. [`🔗 enactic/openarm`](https://github.com/enactic/openarm) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49684289) --- ## 44. Microsoft removes the COPILOT function from Excel — three months after preview hype - **Velocity:** ▮ steady - **Source:** Microsoft Support · confirmed on the official function page - **Tags:** `microsoft` `excel` `ai-features` Quiet but verifiable: Microsoft's own support page now states "Starting September 14, 2026, the COPILOT function is no longer available in Microsoft Excel." The function — which let a cell call an LLM as part of a formula — had been a Frontier/Insider-channel preview; Microsoft points users to the Copilot pane instead, which "supports many of the same AI-powered tasks" but lives outside the grid. No formula-level replacement has been announced. **Why it matters:** Excel's COPILOT() was the flagship demo of AI-as-a-first-class-formula — deterministic, auditable, in-cell. Its removal without a replacement is a real signal about how hard it is to make non-deterministic calls live inside a recalculation engine, whatever the launch-stage messaging said. [`🔗 Microsoft: COPILOT function`](https://support.microsoft.com/en-us/excel/functions/copilot-function) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49706481) --- ## 45. "I tried to hold the company to its ethical commitments" — a DeepMind alumnus goes public - **Velocity:** ▮ steady - **Source:** The Guardian · 36+ pts on HN · 9.8h ago (~10:25 UTC+8) - **Tags:** `ai-safety` `google-deepmind` `essay` Alex Turner's Guardian op-ed adds a DeepMind voice to the week's safety exits and warnings: the former researcher describes trying to hold the company to its ethical commitments against supplying AI for military purposes, and argues the public should take the internal warnings seriously. It runs alongside the Guardian's wider piece on why DeepMind insiders think the public isn't hearing enough. **Why it matters:** The week's safety discourse has run from Bengio's "predictable" agent misbehavior to Cantrill's fear-contagion counterpoint; a named ex-DeepMind account of the internal-advocacy path failing is a different data point — about incentives inside labs, not about model capabilities. It's an op-ed; weight it as testimony, not measurement. [`🔗 The Guardian`](https://www.theguardian.com/technology/2026/sep/14/google-deepmind-ai-warnings) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49706941) --- ## 46. Redis City: an interactive 3D model of everything Redis does — LLM-built frontend, author-corrected framing - **Velocity:** ▮ steady - **Source:** Hacker News (Show HN) · 117+ pts · front page since Sep 12 - **Tags:** `redis` `visualization` `databases` A browsable 3D city where every Redis structure and operation is a building: keys, hashes, streams, pub/sub, expiration, persistence — you can walk the data structures as architecture and watch operations move through them. Part of a small wave of "render the database" explainers that treat systems education as an environment rather than a diagram. **Why it matters:** Corrected in place 09-16: this item originally called the project "hand-crafted to teach, not generated to fill space" — the author says the opposite in the HN thread: "Most of the frontend was built using LLMs… I read the library's code and articles about its architecture, then visualized the concepts using GPT and local models," adding he sees "no point in creating such visualizations manually" (the articles are his own; a bot translates them into three languages). So this is not a counterweight to agent-built work — it is a case study in it: human-held architecture understanding, LLM-executed rendering, with the pedagogy claim resting on the per-component source citations rather than on who typed the meshes. [`🔗 Redis City`](https://poltora.dev/redis) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49676425) --- ## 47. Every invoice in Brazil's economy runs on SOAP 1.2 — and someone catalogued all of it - **Velocity:** ▮ steady - **Source:** Hacker News · 77+ pts · 57h ago (Sep 13) - **Tags:** `integration` `xml` `brazil` The SEFAZ tax-authority webservices — NF-e, NFC-e, CT-e, MDF-e — that carry every fiscal document in Brazil are SOAP 1.2 with mTLS certificates tied to an e-CNPJ, requiring signed XML in portuguese-specific canonical forms. A developer has packaged the complete, production-tested Postman collection of the distribution endpoints, turning tribal integration knowledge into a runnable artifact. The HN thread is a support group. **Why it matters:** The world's most interesting APIs are rarely the trendy ones: an entire nation's fiscal plumbing runs on 2000s-era XML standards with real cryptographic ceremony, and every developer in that economy carries the scars. Documentation-as-collection is also quietly one of the most useful formats an open-source repo can take. [`🔗 stoix-dev/sefaz-webservices-postman`](https://github.com/stoix-dev/sefaz-webservices-postman) · [`🔗 HN discussion`](https://news.ycombinator.com/item?id=49679653) --- ## Metadata | Field | Value | |-------|-------| | Generated | 2026-09-15T12:20:00Z | | Items | 47 | | Sources tracked | 39 (Hacker News, GitHub Trending daily+weekly, CISA KEV, vendor blogs and advisories, EU Commission, Apple Newsroom, courts, security research blogs) | | 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](2026-09-14.md) · [Raw .md](https://trending.md/en/feed/latest.md) · [Archive](../archive/index.md)