1. F-Droid 2.0: the biggest app update in a decade rebuilds the FOSS app store

• Velocity: ▮▮▮ trending
• Source: F-Droid · HN 622+ pts · 187 comments · ~9h ago (~19:30 UTC+8)
• Tags: f-droid android open-source release

F-Droid shipped 2.0 (announced September 24, rolling out over coming weeks after 14
test releases) — the client's largest update in 10 years: a Kotlin/Jetpack Compose
rewrite, a simplified three-area navigation (Discover / Search / My Apps), expanded
categories, and much better CJK search via indexing of descriptions and translated
content. The flagship feature is a unified installer built on Android's new
pre-approval API — enabled partly by EU Digital Markets Act pressure — letting users
confirm installs before download completes, with automatic background update fetching
now the default. The announcement is candid about regressions: Android 6 support
dropped, the Privileged Extension is ignored, panic-triggered app wiping (Ripple) is
temporarily gone, the calculator disguise is simplified, and Nearby sharing isn't in
the initial release. Funded by NLnet's Mobifree fund and OTF; a security audit by
OTF's Security Lab is complete with the report pending.

Why it matters: The EU app-store fight is usually told through commercial stores —
this is the DMA quietly making life easier for the volunteer-run FOSS store too, and
the changelog honestly prices what the rewrite cost in accessibility features.

🔗 F-Droid announcement · 🔗 HN discussion

---

2. Two-tier encryption in the UK: the iCloud customers the government locked out

• Velocity: ▮▮▮ trending
• Source: MacAnorak · HN 320+ pts · 328 comments · ~11h ago (~17:30 UTC+8)
• Tags: encryption privacy uk policy

A detailed explainer of the aftermath of the UK's January 2025 Technical Capability
Notice against Apple: when Apple withdrew Advanced Data Protection for new UK users
rather than build a backdoor, it created two classes of Britons. "Alice," who enabled
ADP before the February 2025 withdrawal, keeps end-to-end encryption across 23 iCloud
categories — because Apple deliberately designed ADP so only the user's own trusted
devices can disable it, servers can't remotely switch it off. "Bill," a new customer,
cannot activate it at all and sits on Standard Data Protection across 14 default-E2EE
categories. The piece traces the fight from Cook's 2014 "there is no back door"
through San Bernardino to the current state: a narrowed UK-only notice, Apple's July
2026 Investigatory Powers Tribunal complaint, and the September 11 Wyden/Davidson
letter urging the IPT to open its gagged proceedings. The author's conclusion: any
mechanism for authorized access is a findable-and-exploitable mechanism, and the
unfairness lands on the government's decision, not Apple's.

Why it matters: The clearest single illustration of what "encrypt-or-backdoor"
standoffs actually produce — not a winner, but a permanent two-tier user base, decided
in secret.

🔗 MacAnorak · 🔗 HN discussion

---

3. A firmware update bricked Samsung smart fridges — and the food spoiled with them

• Velocity: ▮▮▮ trending
• Source: Ars Technica · HN 250+ pts · 244 comments · ~12h ago (~16:30 UTC+8)
• Tags: iot firmware samsung reliability

A SmartThings software update pushed to Samsung smart fridges began bricking units on
the afternoon of September 22, primarily in South Korea — fridges shut down
completely, with spoiled food inside. Samsung attributed the failure to "an error
during its internal testing" and suspended the rollout after reports flooded the
company's Korean community forum. The HN thread (244 comments) turned it into the
week's referendum on appliance connectivity: a refrigerator is a food-safety device
whose failure mode is now "remote code push," and owners of a $2,000+ appliance found
out the hard way that the update channel is a single point of failure. No word yet on
compensation for spoiled food or a timeline for un-bricking affected units.

Why it matters: The strongest recent argument for local-first appliance firmware —
when your fridge requires the cloud to stay cold, every vendor testing mistake becomes
a physical-world loss event.

🔗 Ars Technica · 🔗 HN discussion

---

4. GitHub removed malicious imitation software — 10 minutes after it hit the HN front page

• Velocity: ▮▮▮ trending
• Source: Successful Software · HN 211+ pts · 88 comments · ~7h ago (~21:30 UTC+8)
• Tags: supply-chain github malware trust-safety

Andy Brice's data-wrangling product Easy Data Transform was imitated by a GitHub repo
that copied its name and logo; the fake Mac .dmg triggered multiple malware warnings
on VirusTotal, and the attackers even swapped the disk image's background picture to
tell downloaders to ignore malware alerts. Brice reported it August 31 and got only
an automated reply; he submitted VirusTotal evidence September 10; after 23 days of
silence he published "This is pisspoor. Do better GitHub" — and the repo was removed
within about 10 minutes of his post reaching the HN front page ("Total coincidence.
I'm sure!"). A commenter reports a similar unresolved case since June. The takeaway
the author draws is aimed at users: download software directly from the vendor, and
don't trust the platform to police imitations proactively.

Why it matters: Brand-imitation malware distribution on GitHub is a supply-chain
channel that currently scales with virality, not with harm — and the SLA for removal
appears to be "front page of Hacker News."

🔗 Successful Software · 🔗 HN discussion

---

5. CVE-2026-61732: a CVSS 10.0 prompt-injection RCE in an autonomous red-team agent

• Velocity: ▮▮▮ trending
• Source: NVD / GitHub advisories · CVSS 10.0 (GitHub CNA, NVD Secondary) · published Sept 24, fixed in 1.1.17
• Tags: cve prompt-injection agent-security chatml

Decepticon (BitterSecurity), an autonomous hacking agent for red teams, wrapped web
crawl results — the output of its own reconnaissance — into LLM messages without
neutralizing ChatML special-token literals. Under the BYOK model (user-supplied keys
to any OpenAI-compatible endpoint), most self-hosted inference servers (vLLM, SGLang,
Ollama, LM Studio, text-generation-webui) don't filter special-token literals from
user content by default — so those literals parse as structural role-boundary tokens,
and a string planted in a target web page forges a new operator turn the model
treats as authoritative, yielding arbitrary command execution inside the agent's Kali
sandbox. Fixed in 1.1.17 (fix commit and GHSA-g5f9-3xfg-p9mf referenced on NVD).
CISA-coordinated SSVC assessment on the record: exploitation "PoC," automatable
"yes," technical impact "total."

Why it matters: The prompt-injection-to-shell pipeline, formally — an offensive
agent CVE that is simultaneously a warning for every agentic tool that feeds web
content to a self-hosted model with unfiltered special tokens.

🔗 NVD record · 🔗 GHSA-g5f9-3xfg-p9mf

---

6. The newest ESP32 can actually run Linux — and it's getting close to a Raspberry Pi

• Velocity: ▮▮ rising
• Source: XDA Developers · HN 195+ pts · 95 comments · ~8h ago (~20:30 UTC+8)
• Tags: esp32 hardware linux riscv

The ESP32-P4 generation can natively boot Linux — a first for the family that
previously only emulated it in limited ways — via its dual-core RISC-V cores, with
Wi-Fi 6 + Bluetooth 5.4 coming from the paired ESP32-C5 companion chip, Thread/Zigbee,
Ethernet, USB 2.0 OTG and 32–64 MB PSRAM configurations. The article's emphasis is on
peripheral support: Pi-compatible CSI camera connectors, microSD, and projects like
ESP-KVM that already use the P4 as a Pi-class component. Fair framing from the HN
thread: this is a microcontroller becoming SBC-adjacent, not a Pi replacement — RAM
ceilings and GPU-class workloads remain Pi territory — but the "ESP32 cannot run
Linux" era is officially over.

Why it matters: The $5-ish RISC-V microcontroller line crossing the "boots
Linux" line compresses the floor of what a networked, Linux-capable device costs.

🔗 XDA Developers · 🔗 HN discussion

---

7. Bastardica: a foundry for "cursed" hybrid fonts that run on OpenType ligatures

• Velocity: ▮▮ rising
• Source: Show HN · 185+ pts · 32 comments · ~14h ago (~14:30 UTC+8)
• Tags: fonts typography web pyodide

A browser tool for making "bastard" fonts in the style of Times New Bastard: pick a
base font plus mix-in fonts, and generated output replaces chosen glyphs (the
canonical example: everything after "Th" comes from Comic Sans). The clever part is
the delivery mechanism — the glyph swap is implemented as a liga contextual
substitution registered for every script, so browsers enable it by default and the
font works anywhere text is shaped, including design tools and print. Everything runs
locally via Pyodide (Python in WASM) and fontTools — no fonts are uploaded — and
exports to TTF/OTF/WOFF2. Documented limits: apps that disable ligatures won't show
the mix, 3+ fonts can collide on shared intervals (use prime-number strides), and the
licensing note is honest — a mixed font is a derivative work of every source font.

Why it matters: A delightful bit of font engineering whose real lesson is that
OpenType contextual substitution is an underused client-side rendering primitive.

🔗 Bastardica · 🔗 HN discussion

---

8. Show HN: Whiteboard (YC W26) — an open-source IDE where humans and agents architect software together

• Velocity: ▮▮ rising
• Source: Show HN · 114+ pts · 35 comments · ~4h ago (~00:30 UTC+8)
• Tags: ide agents developer-tools yc-w26

Whiteboard is an MIT-licensed desktop IDE built as a vendored Code - OSS fork (the
authors say ~45% of stock VS Code is Copilot code they don't need) where agents like
Claude Code and Codex get an SDK to draw flowcharts, sequence diagrams and ER diagrams
on a canvas next to the code they describe. Diagram elements and agent-trace quotes
jump to the underlying code; a Rust-based AST-aware semantic diff viewer renders large
new functions as pseudocode and collapses tests; a decision log lets agents link their
own traces so requirements, implementations and autonomous decisions stay inspectable.
Honest limitation list in the README: file editing isn't supported yet, multi-repo
review is weak, and post-share updates don't propagate. Anonymous telemetry excludes
code, diffs, prompts and model output, and can be disabled.

Why it matters: The agent-IDE frontier is moving from "agents edit files" to
"agents and humans share a design surface" — and this one ships the diff and trace
plumbing, not just a canvas.

🔗 devdotfast/whiteboard · 🔗 HN discussion

---

9. Fearless SIMD 1.0: eight years of work to take unsafe out of SIMD

• Velocity: ▮▮ rising
• Source: Linebender · HN 89+ pts · 15 comments · ~2h ago (~02:00 UTC+8)
• Tags: rust simd graphics release

Linebender's fearless_simd reaches 1.0 on stable Rust: portable SIMD abstractions
with function multiversioning via a #[simd] macro, safe access to platform
intrinsics with zero overhead, and zero ad-hoc unsafe blocks — the crate rests on
two audited primitives (the kernel! macro using target feature 1.1, and a
bytemuck/zerocopy-inspired safe transmute module). It offers both
platform-consistent "precise" and platform-fast variants for edge-case-sensitive ops,
plus hardware-native vector widths the standard library won't cover. The team
contributed upstream optimizations to Rust and LLVM along the way; 30 crates use it
directly, over a thousand indirectly. Commitments: 3 years of security updates, and a
stated path to f16, Arm SVE and RISC-V Vector extensions without breaking the API.

Why it matters: SIMD in safe Rust has been a decade-long gap between C++ and
Rust graphics/audio stacks — a 1.0 with a stability promise is what adoption waits
for.

🔗 Linebender blog · 🔗 HN discussion

---

10. "Schrödinger's Code Repository": SWE-bench agents lean on memorized repos

• Velocity: ▮▮ rising
• Source: arXiv / Hugging Face daily papers · ~70 upvotes, trending #1 today
• Tags: benchmark swe-bench memorization evaluation

A Shanghai Jiao Tong University paper (arXiv:2609.27891) asks whether coding agents
actually learned SWE-bench or memorized it. SchrodingerRepo treats each test
repository as an evaluation-time latent variable, applying four progressive,
behavior-preserving transformations: problem-statement reconstruction, namespace
remapping, intra-file layout reordering, and functionality-preserving code rewriting
— keeping executable behavior identical while stripping recognizable cues. Result
across models on SWE-bench Verified and SWE-QA: removing repository cues consistently
degrades performance and substantially increases interaction costs, with the cost
concentrated in repository exploration and localization. The abstract is qualitative
— no headline percentage drop is claimed — and the authors frame the conclusion
carefully: agents "partially rely on memorized repository-side cues."

Why it matters: The memorization critique of SWE-bench now has a clean causal
method — and the honest framing (no inflated single number) makes it harder to
dismiss.

🔗 arXiv:2609.27891 · 🔗 Hugging Face daily papers

---

11. WaveDigger: a web UI for querying Apple's undocumented Wi-Fi location service

• Velocity: ▮ steady
• Source: Show HN · 67+ pts · 10 comments · ~28h ago (~00:30 UTC+8)
• Tags: wifi geolocation privacy nextjs

WaveDigger locates Wi-Fi access points by BSSID and cell towers (LTE and 5G NR) on an
interactive map by querying Apple's undocumented WPS endpoint
(gs-loc.apple.com/clls/wloc) directly with protobuf-encoded requests — including
the China-region variant — building on the reverse-engineering work of the
apple-corelocation-experiments project (protobuf definitions, coordinate encoding,
the required bytes prefix). TypeScript/Next.js 15 with deck.gl, AGPL-3.0, with the
AGPL network-service clause pointedly relevant given what the tool exposes. The
privacy implication is the story: Apple's crowdsourced location database is
effectively a public BSSID search engine, queryable by anyone who implements the
protocol.

Why it matters: Google's WPS has long been a known geolocation oracle; a polished,
self-hostable Apple equivalent makes the "MAC address → physical location" capability
one fork away from any attacker.

🔗 christianrowlands/wavedigger · 🔗 HN discussion

---

12. When the Debugger Lies: J-Link's memory cache vs. a DMA-style peripheral

• Velocity: ▮ steady
• Source: Daniel Mangum · HN 61+ pts · 17 comments · ~30h ago (~22:30 UTC+8)
• Tags: embedded debugging jlink hardware

While provisioning keys into Nordic nRF54L15's Key Management Unit through a SEGGER
J-Link debugger, Mangum found reads kept returning stale values — even after erasing
the device, only the first key push appeared in memory. The root cause: the KMU writes
to RAM independently of the halted CPU (DMA-like), while JLinkGDBServer's memory cache
(SetEnableMemCache, on by default) is never invalidated because the core never
advances. Direct ReadMemAP requests through the AHB-AP access port bypass the cache
and show fresh data; disabling the cache via monitor exec SetEnableMemCache = 0 or
single-stepping the core fixes it. Mangum's diagnosis of why this bites rarely: stale
cache only matters when a peripheral can modify memory while the core is halted —
exactly the configuration security work on MCUs keeps creating.

Why it matters: A debugging-war-story with a transferable rule: hardware that
writes memory behind the CPU's back will defeat any debugger that caches through the
CPU's view.

🔗 When the Debugger Lies · 🔗 HN discussion

---

13. FxEmbed re-trends at +165 stars/day: the link-preview fixer for X and Bluesky

• Velocity: ▮ steady
• Source: GitHub Trending (daily) · 5.3k★ · +165 today
• Tags: embed bluesky twitter cloudflare-workers

FxEmbed — the home of FxTwitter, FixupX and FxBluesky — fixes broken link previews
for X and Bluesky posts in Discord, Telegram and other clients: real video, multiple
images, polls, quote posts and translations, via prefix domains (fxtwitter.com,
fixupx.com, fx + bsky.app). TypeScript on a Cloudflare Worker with a Docker
self-host path, MIT-licensed, 4,207 commits and active CI. The re-trend rides the
continuing decay of X's own embed support — every client-side embed regression sends
another wave of users to prefix domains. The repo notes it is not affiliated with
X Corp, which has been tightening third-party access to embeddable data.

Why it matters: Community infrastructure quietly doing the work a platform
abandoned — with a self-hostable escape hatch if the hosted instance ever disappears.

🔗 FxEmbed/FxEmbed · 🔗 GitHub daily trending

---

14. A web-based IBM 1620 emulator — and software recovery for a 1963 decimal computer

• Velocity: ▮ steady
• Source: Show HN · 42+ pts · 11 comments · ~34h ago (~18:30 UTC+8)
• Tags: emulation retrocomputing ibm web

pkimpel/retro-1620 emulates the IBM 1620 Model 2 — a 1960s transistorized,
decimal, variable-field-length machine with no software-accessible registers and
12-digit instructions — in the browser, with an operating environment around it
(SPS assembler and FORTRAN II era, Monitor batch system, card reader/punch, 1443
line printer, 1311 disk). The Model 2 ran a 10µs memory cycle with hardware
multiply via in-memory lookup tables; about 2,000 units of both models were ever
built, supported until 1970. MIT-licensed, with software recovery sourced from
bitsavers. Small project (17 stars, 170 commits) — the HN post is the vehicle for a
genuinely working piece of computing-history preservation.

Why it matters: Decimal, variable-field-length architecture is a lineage modern
developers never touch — a browser-runnable emulator is the most accessible window
into it that exists.

🔗 pkimpel/retro-1620 · 🔗 HN discussion

---

15. Forging 1024-bit RSA signatures in nearly SNFS time — without ever factoring the key

• Velocity: ▮ steady
• Source: IACR ePrint · HN 30+ pts · 2 comments · ~10h ago (~18:30 UTC+8)
• Tags: cryptography rsa hsm research

Laura Shea, Miro Haller, Adam Suhl, Nadia Heninger (UC San Diego) and Emmanuel Thomé
(INRIA) implemented a 2007 Joux–Naccache–Thomé attack end-to-end against a real HSM:
with temporary access to a raw RSA signing/decryption oracle, an attacker can forge
signatures for any message without ever factoring the key. The 1024-bit demonstration
took 1,380 CPU core-years over five calendar months and 2³² oracle queries; after
precomputation, chosen signatures forge offline in 180 core-years. Extrapolation to
1024–4096 bit keys: RSA's concrete security under this model is 15–30 bits below
standard factoring-based estimates — even 4096-bit RSA may fall short of 128-bit
security here. The scope conditions are explicit and load-bearing: this requires raw
oracle access (an HSM API, a blind signing service) and says nothing about
factoring-only RSA; the large-key numbers are extrapolations. The authors' punchline
is migration guidance: another reason to move off RSA during the post-quantum
transition.

Why it matters: The threat model — forge through the API, never touch the key —
is exactly how HSM-backed PKIs actually expose RSA, and the security-margin haircut
is large enough to matter for long-lived signatures.

🔗 ePrint 2026/2131 · 🔗 HN discussion

---

16. Project Suncatcher's first TPU satellite is days from orbit — Google's ML-in-space moonshot gets a launch date

• Velocity: ▮▮▮ trending
• Source: Google Research blog · HN 139+ pts · 254 comments · ~14h ago (~21:53 UTC+8)
• Tags: google space ml-infrastructure tpu

Google's Project Suncatcher — the moonshot exploring whether scalable ML compute can
live in low Earth orbit, where satellites can capture "up to eight times more solar
power than on Earth" — is days from its first hardware test: a prototype satellite
launching on SpaceX's upcoming Transporter-18 rideshare mission, developed with
Planet. The engineering writeup is unusually concrete about what got tested: launch
loads of ~10 g sustained (chips see 50–100 g) survived three-axis vibration testing;
Trillium TPUs bombarded with protons at UC Davis's Crocker Nuclear Laboratory while
running workloads held up past the total ionizing dose of a five-year mission; and
vacuum cooling via heat pipes plus radiators has so far only been validated in a
thermal vacuum chamber. The 2027 milestone is two satellites testing high-bandwidth
short-range laser links. The team's own caveats are blunt: "some things can only be
tested in space," this launch is exploratory rather than proof, and today's laser
systems are built for exactly the opposite regime (low bandwidth, long distance).

Why it matters: The post prices the gap between "satellites with TPUs" and
"datacenters in orbit" — radiation and launch turn out survivable, while cooling and
laser interconnects are the unproven load-bearing parts.

🔗 Google Research blog · 🔗 HN discussion

---

17. SourceHut account takeover: ANSI escape sequences in CI build logs rendered as live XSS for 4.5 years

• Velocity: ▮▮▮ trending
• Source: Arusekk · HN 90+ pts · 13 comments · ~8h ago (~03:54 UTC+8)
• Tags: security xss ci ansi2html

Arusekk's writeup traces a bug (CVE-2026-92973) in ansi2html, the library
builds.sr.ht uses to render CI logs: it converted ANSI OSC 8 hyperlink sequences to
<a> tags without breaking out of the attribute correctly and without blocking
javascript: URLs, so an attacker could plant onfocus= handlers that execute in
the session of anyone viewing the job page. Injection required no account — a patch
to a public mailing list with CI enabled, or any remote resource printed to a log,
was enough. Because the page carries the victim's CSRF token and builds.sr.ht holds
deploy keys for sr.ht itself, the author grades it as account takeover and
wormable. The flaw persisted roughly 4.5 years; builds.sr.ht shipped an
auto-sanitizing workaround August 4, and the real fix landed in ansi2html 1.9.4 on
September 2. A telling footnote: the author proposed a CVSS 4.0 vector arguing
"high or critical, not just a mere medium" — and is openly annoyed that VulnCheck
altered it.

Why it matters: Untrusted text arriving through a rendering layer is the oldest
XSS shape there is — and CI logs are attacker-writable input on nearly every forge.

🔗 blog.arusekk.pl · 🔗 HN discussion

---

18. browser-use ships "jev-ultrafast": a web agent whose every observation is an indexed action space — 19.9k★ in nine days

• Velocity: ▮▮▮ trending
• Source: GitHub · browser-use/jev-ultrafast · 19.9k★ · created Sept 16, pushed today
• Tags: agents browser-automation jev system-1

browser-use's new repo pairs TypeSafe's Jev decision model with a reworked agent
loop: each page observation becomes a numbered element table, and one Jev request
picks both an operation (CLICK / TYPE_TEXT / SELECT / SCROLL / WAIT / DONE / BLOCKED)
and a target — target heads only contain compatible elements, so the two decisions
cost a single network round trip. Text is generated by a small helper model only
when TYPE_TEXT is chosen (the recorded demo uses inception/mercury-2.5 with
reasoning disabled). The repo's own measurements are unusually honest: the flagship
Google Flights demo verifies at 7.073 s, and the matched six-run comparison shows
median 9.45 s → 7.09 s (25% lower, 1,092 → 101 browser protocol calls) — but with
three pairs, the README itself notes "three pairs are too few for a strong
statistical claim (two-sided sign-test p = 0.25)" and that this is "a small
controlled-input comparison, not a broad agent benchmark." MIT-licensed, with a
Browser Use Cloud waitlist attached. One caveat this feed's own star-integrity
check (Sept 26) adds to the momentum in the title: those stars sit on just three
visible main-branch commits (~6,806★/commit — squash-dropped main, with
development living on seven unmerged codex/* branches), so read the ★ as
attention, not visible engineering.

Why it matters: The System-1 decision-model pattern (one scored choice per
round trip, text generation only when needed) is becoming a real agent-runtime
architecture — and this repo publishes its own weak statistics next to its
headline number, which is rarer than the speed. The star momentum, though, is
a separate claim from the honest-stats one: three visible main-branch commits
mean the ★ count measures attention, not engineering history.

🔗 browser-use/jev-ultrafast · 🔗 Performance measurements

---

19. GitLab's rough night: a 2.5-hour GitLab.com outage lands the same day as patch releases fixing two CVSS 9.9 RCEs

• Velocity: ▮▮ rising
• Source: status.gitlab.com · HN 58+ pts · 28 comments · ~5h ago (~07:10 UTC+8)
• Tags: gitlab outage security-release cve

GitLab.com began returning 503s at 23:02 UTC on September 24 across the website,
API, Git operations, both registries, Pages, CI/CD runners and SAML SSO; the cause
was identified 25 minutes in, and all 23 components were back to operational by
03:36 UTC September 25 — roughly a 2.5-hour disruption, now in monitoring. The same
day, GitLab's patch releases (19.2.7 / 19.3.3 / 19.4.1) fixed two CVSS 9.9
authenticated remote-code-execution issues in CI/CD configuration parsing: a double
free (CVE-2026-89078) and an integer overflow (CVE-2026-93577), both triggered by
specially crafted regular expressions, both via HackerOne, both scored by GitLab as
CNA. Unrelated in mechanism — but a self-hosted admin patching tonight and an
SaaS user staring at 503s got the same message about their forge's week.

Why it matters: Two CVSS 9.9s reachable by any authenticated user through CI
config is upgrade-now territory for every self-hosted GitLab — the outage just
guaranteed the release gets read.

🔗 NVD: CVE-2026-89078 · 🔗 NVD: CVE-2026-93577 · 🔗 GitLab status

---

20. Japanese used bookstores see a 5× sales surge — as buyers ship books abroad "by the ton" for suspected AI scan-and-shred

• Velocity: ▮▮ rising
• Source: Tom's Hardware · HN 77+ pts · 122 comments · ~13h ago (~22:51 UTC+8)
• Tags: ai-training-data publishing copyright japan

Tom's Hardware reports on a striking side effect of the race for training data:
Japanese used bookstores seeing roughly 5× normal sales, driven not by readers but
by unidentified bulk buyers purchasing "by the ton" — including one 50-ton order
shipped to the US, suspected of ending up in facilities that scan books for AI
training and then destroy them. The AI destination remains suspected rather than
confirmed, and the report frames the buyers as anonymous; what is documented is the
scale and the pattern of purchases, and the cultural concern in Japan about printed
works being physically consumed as dataset feedstock. The HN thread immediately
connected it to the archival-access bottleneck that even AI-for-research advocates
keep flagging: the manuscripts most worth training on are the ones nobody has
digitized — and this pipeline digitizes and destroys in one step.

Why it matters: Physical books are becoming a scraping target — with the
"scrape" being a shredder — a data-acquisition story that's also a preservation
story.

🔗 Tom's Hardware · 🔗 HN discussion

---

21. "AI labs need to start funding historical research": a historian runs frontier models on 17th-century archives — and writes down what breaks

• Velocity: ▮▮ rising
• Source: Res Obscura (Benjamin Breen) · HN 97+ pts · 14 comments · ~9h ago (~03:14 UTC+8)
• Tags: llm digital-humanities research history

Historian Benjamin Breen's essay argues that frontier models have crossed a
threshold for archival research, with receipts: an Opus agent downloaded 5,000+
files from Samuel Hartlib's digitized archive and spawned sub-agents to search
Google Books; GPT-6 Astra concluded John Dee's Liber Loagaeth is mostly nonsense
syllables rather than code — and showed via character-frequency analysis that Edward
Kelley got lazier over time, cross-checked against Dee's diary; and a possibly-new
finding that Newton and Hartlib used different anagrams for the same substance
(Hungarian vitriol) with matching quantities. Breen keeps the score honest: the
Charles V cipher letters Opus partially deciphered had already been solved (one in
the 1530s, one in 1916) — "desk research" the model skipped — the Newton finding
only "seems" new, and the real bottleneck is that most premodern manuscripts are
undigitized. His three proposals: digitize and open the archives, give historians
free compute, and let historians nominate solvable "millennium problems."

Why it matters: A working researcher's measured account of where LLM agents
genuinely add archival value — and of the already-solved-problem failures that
vendors' demos never mention.

🔗 Res Obscura · 🔗 HN discussion

---

22. "Dynamic Abliteration": runtime refusal suppression with frozen weights — a proof of concept that admits what it is

• Velocity: ▮▮ rising
• Source: Madhukar Anand (Solvy Tech blog) · HN 105+ pts · 40 comments · ~13h ago (~22:33 UTC+8)
• Tags: alignment llm steering open-weights

Instead of abliteration's permanent weight edits, this post steers refusal behavior
at inference time: PyTorch forward hooks intercept residual streams at layers
12–20 of Qwen3-4B and add gate × projection(n-gram memory), where an
Engram-inspired module — a dynamic sigmoid gate, an O(1) four-table n-gram hash,
and learned per-layer projections — only fires when refusal triggers appear. It's
trained on 2,000 PKU-SafeRLHF samples in 8.92 minutes on one A100, with base
weights frozen. The post is candid about the failure that motivates it: single-layer
steering at alpha 1.2 still produces refusals because downstream layers reconstruct
the behavior. What it does not evaluate: anything about misuse — the author states
the steered model will comply with harmful requests, presents the whole thing as a
proof of concept on one model, and notes the code was AI-generated.

Why it matters: Runtime steering with frozen weights is the deployment-friendly
version of a technique that's usually discussed as vandalism — and the post's
candor about statistical and safety limits makes it a usable reference point.

🔗 Solvy Tech blog · 🔗 HN discussion

---

23. Fire at a Starlink ground station near Warsaw treated as suspected arson targeting critical infrastructure

• Velocity: ▮▮ rising
• Source: Notes from Poland · HN 167+ pts · 172 comments · ~18h ago (~17:52 UTC+8)
• Tags: starlink infrastructure sabotage poland

Firefighters responded to a blaze around 9 p.m. Wednesday at a ground station in
Wola Krobowska, south of Warsaw — a facility owned by Polish state telecom Exatel
that, per digital affairs minister Krzysztof Gawkowski, "provides internet
transmission through Poland, and also to Ukraine" and serves Central and Eastern
Europe alongside a station in Lithuania. Police and the Internal Security Agency
were dispatched on suspicion of deliberate ignition; Gawkowski called it "an arson
attack that was intended to impact critical telecommunications infrastructure" with
a "modus operandi... clearly Russian" — while other officials stayed careful:
security-services spokesman Jacek Dobrzyński said it was "still too early to talk
about the causes or motives." The attribution hedging is the story's honest part;
the strategic fact is not: Starlink's terrestrial footprint in a war zone is now
being physically attacked, and Ukraine's connectivity runs through Polish
buildings.

Why it matters: Satellite internet's dependence on ground stations keeps
turning "space infrastructure" into ordinary, flammable, politically legible
real estate.

🔗 Notes from Poland · 🔗 HN discussion

---

24. mammoth.js: prototype pollution in .docx styles chains to local-file disclosure — fixed in 1.12.2

• Velocity: ▮▮ rising
• Source: NVD · CVE-2026-97151 · CVSS 8.4 (CVSS 4.0, MITRE CNA) · published Sept 24
• Tags: cve prototype-pollution nodejs docx

mammoth, the widely used docx→HTML converter, allowed arbitrary properties on
Object.prototype when reading style definitions from a crafted document. The
sharper part is the chain: in 1.11.0 through 1.12.1, applications that convert
multiple documents in the same process and return the converted HTML could disclose
the contents of local server files to whoever supplies the documents, by polluting
toward externalFileAccess: true. Fixed in 1.12.2 (two commits referenced on NVD).
The affected pattern — a server converting user-uploaded documents into HTML it
then serves back — is one of the most common document-processing shapes on the
web, and prototype pollution is exactly the bug class that survives framework-level
sanitization because it happens below the template layer.

Why it matters: "Just convert the docx" is load-bearing infrastructure for
document pipelines everywhere; this is the second reminder this month that the
conversion step is an attack surface, not a utility.

🔗 NVD record · 🔗 mwilliamson/mammoth.js

---

25. SigNoz: an empty default JWT secret let anyone forge admin sessions — including non-revocable 30-day refresh tokens

• Velocity: ▮ rising
• Source: NVD / VulnCheck · CVE-2026-97055 · CVSS 9.2 (CVSS 4.0, VulnCheck) / 8.1 (v3.1) · published Sept 24, fixed in v0.143.0
• Tags: cve signoz jwt observability

SigNoz, the open-source observability platform, shipped with its JWT tokenizer
signing secret (SIGNOZ_TOKENIZER_JWT_SECRET) defaulting to an empty string — and
Config.Validate() didn't reject it, so unconfigured deployments signed and
verified session tokens with an empty HMAC key. Since the JWT provider was the
default, any such deployment was exposed: an unauthenticated attacker who knows an
existing user's ID (obtainable without auth from /api/v2/sessions/context, along
with the org ID) could forge valid tokens for that user, including administrators.
Forged refresh tokens exchanged at /api/v2/sessions/rotate cannot be revoked and
last 30 days by default. v0.143.0 requires a secret when the jwt provider is
selected and switches the default to opaque tokens. An observability platform is
the worst possible place for this class of bug: it holds every log, trace and
metric your stack emits.

Why it matters: "Empty default credential" bugs keep landing in
self-hostable infrastructure — and a forgery primitive that yields non-revocable
30-day sessions converts an observability tool into persistent access.

🔗 NVD record · 🔗 Fix commit

---

26. m3e-canvas: sketch Material 3 Expressive screens in the browser, hand them to your coding agent as prompts — 8.2k★ and Trendshift #1

• Velocity: ▮ rising
• Source: GitHub Trending · lnkiai/m3e-canvas · 8.2k★ · +Trendshift #1 of the day
• Tags: material-design design-tools generative-ui vibe-coding

m3e-canvas is a drag-and-drop editor for Material 3 Expressive UI screens that runs
entirely in the browser (Next.js static export, React 19, everything in
localStorage, no backend): multi-screen phone and desktop flows with tap/swipe
transitions, theming across M3E's four axes (color, shape, typography, motion), and
magnetic connections. The design decision doing the work: it deliberately does not
generate code — it emits natural-language prompts (in Japanese, English, Chinese or
Korean, targeting Android or web) that you paste into Claude Code, Codex, Gemini
CLI or Cursor. An optional AI helper uses your own API key, talking straight to the
provider. Documented limits: mobile editing is single-screen and buttons-only, and
share links plus AI-assisted drafting are beta.

Why it matters: The design-to-agent handoff is settling on prompts, not code —
the canvas owns intent, and the coding agent owns implementation.

🔗 lnkiai/m3e-canvas · 🔗 Live demo

---

27. Compositor: a free, open-source "Photoshop alternative for Mac" ships three releases in three days

• Velocity: ▮ rising
• Source: GitHub · robbietilton/Compositor · 5.4k★ · v1.2.9–v1.2.11 released Sept 23–24
• Tags: macos image-editing open-source native-apps

Compositor is a native macOS image editor (MIT, requires macOS 26.5+) built for
compositing and post-processing: layers with GPU-rendered effects, clipping masks,
adjustment layers, Photoshop's full blend-mode set, non-destructive transforms,
retouch tools (Spot Healing, Clone Stamp, Content-Aware Fill), a Camera Raw filter,
and PSD/PSB import with Photoshop-remappable shortcuts. The author's stated motive:
Photoshop's price, and GIMP disrupting his compositing workflow. The release pace —
three signed, notarized DMG releases in three days — matters as much as the feature
list. The README is honest about import fidelity: PSD/PSB support is 8-bit RGB only
(explicitly not CMYK), and vectors and vertical text become pixels on import, with
only folders, masks, blend modes and simple horizontal text staying editable.

Why it matters: A credible native Photoshop-shaped editor is the last big hole
in the open-source creative stack — the PSD round-trip caveats are exactly what
will decide whether it converts working professionals.

🔗 robbietilton/Compositor · 🔗 Releases

---

28. Search: a ~3 MB WebKit browser for macOS with no toolbar, no accounts, and no telemetry

• Velocity: ▮ steady
• Source: Show HN · 62+ pts · 24 comments · ~19h ago (~17:11 UTC+8)
• Tags: browser macos webkit minimalism

Office Commun's Search (MIT, ~12,700 lines of Swift, no third-party dependencies)
is a deliberately minimal browser on macOS's system WebKit: one address/search
field, tabs, and nothing else — no toolbar, start page, sidebar, accounts or sync.
The engineering is more substantive than the concept suggests: a network-level ad
blocker via WKContentRuleList that runs before requests fire, lazy tab
restoration (reopened tabs cost nothing until clicked), persistent per-site element
hiding, private tabs with separate cookie jars, and Chrome extension support
through WebKit's own extension engine with API shims. Passwords stay in the macOS
keychain, with one-click import from Chrome, Arc, Brave and Edge. The only outbound
requests are page loads and one daily update check.

Why it matters: As browsers accrete agents, workspaces and AI sidebars, a
3 MB browser that does the 2010 core feature set natively is a useful statement
about how much of the modern browser is optional.

🔗 driceroland/Search · 🔗 HN discussion

---

29. Best LLM for every budget: a daily-refreshed price/intelligence frontier chart, honest about its own caveats

• Velocity: ▮ steady
• Source: Show HN · 167+ pts · 105 comments · ~14h ago (~22:09 UTC+8)
• Tags: llm benchmarks pricing data-visualization

terryds/bestvaluemodel plots every model from Artificial Analysis's Intelligence
Index against its blended API price (3:1 input:output, per 1M tokens) on a
log-scale chart, refreshed daily by a GitHub Actions cron with a "what changed"
diff between snapshots. The core object is the value frontier: models where
"nothing cheaper is also smarter," computed by keeping each model that scores
higher than everything cheaper, with one row per model (best effort variant only)
and a min-score filter so an ultra-cheap dumb model can't anchor the line. The
methodology page states its own limits: cached-input discounts, batch pricing and
fast modes are excluded, and the Index is re-based between versions so scores only
compare within a snapshot.

Why it matters: Most LLM leaderboards optimize for "what's best"; this one
answers the question people actually budget against — and shows its frontier math
instead of a ranking vibe.

🔗 bestmodelforyourbudget · 🔗 terryds/bestvaluemodel

---

30. DAWO: the Dutch government is building a Microsoft-alternative workplace on NixOS

• Velocity: ▮▮▮ trending
• Source: HN 386+ pts · 173 comments · ~4h ago (~16:06 UTC+8)
• Tags: nixos government open-source digital-sovereignty

DAWO is the Dutch government's open community for a "digitally autonomous workplace"
— a shared blueprint rather than one product, organized into four inspectable,
replaceable building-block categories: AI, operating system (DAWO-NixOS, "installation
blocks for a reproducible workplace"), cloud, and collaboration software. It is led
from MinBZK (the Ministry of the Interior and Kingdom Relations), with code public on
code.overheid.nl and Codeberg, and five stated goals headed by digital autonomy and
verifiability of government IT. The site is candid about being early: no roadmap,
budget or adoption milestones are published — the visible activity is working
sessions, docs and pilots.

Why it matters: Europe's digital-sovereignty wave keeps producing "we'll switch
offices" announcements; this one picks NixOS specifically, making reproducibility —
the property that makes a government desktop auditable — the load-bearing
architectural choice.

🔗 dawo.community · 🔗 HN discussion

---

31. Meta took down the AI-glasses video filmed at Meta — for "bullying and harassment"

• Velocity: ▮▮▮ trending
• Source: Reddit via HN · 614+ pts · ~28h ago (~16:23 UTC+8)
• Tags: meta ai-glasses content-moderation privacy

Dutch satirical creator Roel Maalderink, working with digital-rights organization
Bits of Freedom, visited Meta's Amsterdam office and filmed Meta's own employees with
Meta's own camera glasses. The staff's discomfort at being recorded was the point of
the piece: the hardware's recording indicator exists precisely so people know when
they're being filmed, and the video is the demonstration of what that feels like.
Meta removed the video from Facebook and Instagram citing its bullying-and-harassment
policy; the YouTube version now carries the title "Meta's staff hated Meta's glasses.
Meta deleted the video." HN's thread turned into the consent debate the video was
staging — with commenters split on whether public-place recording of unwilling
subjects is hypocrisy or journalism.

Why it matters: The moderation rule built to protect users from harassment was
applied to suppress criticism of the company's own product — and the suppressed
argument (recording indicators don't make surveillance comfortable) is exactly the
one Meta's glasses business needs to answer.

🔗 HN discussion · 🔗 "Meta's staff hated Meta's glasses. Meta deleted the video." (YouTube)

---

32. Rails World 2026 keynote: Hey's email backend rewritten in Rust by agents — and DHH says he's retired from programming

• Velocity: ▮▮ rising
• Source: HN 356+ pts · 387 comments · ~45h ago (~23:33 UTC+8, Sep 23)
• Tags: rails rust agentic-coding keynote

The Rails World 2026 opening keynote contained no Rails roadmap — commenters
repeatedly noted "Rails content not found" — and that absence frames what DHH chose
to announce instead: Hey.com's email backend has been reimplemented in Rust using
agentic coding, with a claimed ~90% reduction in server deployment; his team shipped
and maintains six native mobile apps with agents (after trying UX staff as the
vibe-coders and watching it fail — engineers did the coding); and he stated "I have
retired from being a professional programmer." The closing framing was p(bloom)
versus p(doom), with Rails positioned as well-suited to agentic coding — the
framework now publishes its own AI model benchmarks at rubyonrails.org/ai.

Why it matters: The most AI-bullish mainstream framework founder's flagship
concrete result is a production rewrite in a different language — and his own team's
failed "non-engineers prompt it" experiment is a rare first-hand negative datapoint
in the agent-productivity literature.

🔗 Rails World 2026 Opening Keynote (video) · 🔗 HN discussion

---

33. CVE-2025-13032, part 2: Avast's kernel driver, from double-fetch to SYSTEM

• Velocity: ▮▮ rising
• Source: SAFA Team · HN 66+ pts · ~5h ago (~15:03 UTC+8)
• Tags: cve avast kernel exploitation

SAFA Team published the final part of its Avast Antivirus research: a complete
exploitation walkthrough of CVE-2025-13032 — a double-fetch in Avast's sandbox kernel
driver (affecting Avast/AVG before 25.3 on Windows) — on an up-to-date Windows 11
system. The chain: the controlled paged-pool overflow becomes arbitrary kernel
read/write by corrupting the RegBuffers array of an IORing object, via heap spray,
kernel-address leak through MDL introspection, deliberate repairs to avoid a blue
screen on teardown, and finally SYSTEM via token theft. The score carries a live
disagreement worth recording: 9.9 critical from the vendor CNA (Gen Digital)
versus 7.8 high from NVD.

Why it matters: Endpoint-security kernel drivers are themselves privileged attack
surface, and a full modern-Windows exploit chain published end-to-end — including the
BSOD-avoidance plumbing most writeups skip — is reference material for both sides.

🔗 SAFA Team research · 🔗 NVD: CVE-2025-13032

---

34. Adobe Commerce/Magento CVE-2026-71362 lands on CISA KEV — CVSS 9.1 incorrect authorization, no user interaction

• Velocity: ▮▮ rising
• Source: NVD / CISA KEV · CVSS 9.1 (Adobe PSIRT CNA, NVD Analyzed) · KEV added Sept 24
• Tags: cve kev magento adobe

CVE-2026-71362 is an incorrect-authorization flaw in Adobe Commerce and Magento that
yields privilege escalation — an attacker can gain elevated access to sensitive
resources with no user interaction required. Adobe's own PSIRT scored it CVSS 9.1
critical (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:N), NVD analyzed and corroborated, and
the fix ships in Adobe's APSB26-92 advisory. The KEV addition on September 24 puts
it in exploited-in-the-wild territory, which for a Commerce deployment means the
federal patch clock and the practical assumption of active targeting.

Why it matters: Authorization bugs in commerce backends sit directly on payment
and customer data flows — and a KEV listing is the signal that someone besides the
researcher is using it.

🔗 NVD: CVE-2026-71362 · 🔗 Adobe APSB26-92

---

35. Paperclip: "if OpenClaw is an employee, Paperclip is the company" — agent orchestration hits trending #1 at 83.4k★

• Velocity: ▮▮▮ trending
• Source: GitHub · paperclipai/paperclip · 83.4k★ · trending #1 · v2026.916.1 released Sept 21
• Tags: agents orchestration nodejs open-source

Paperclip is an MIT-licensed Node.js server + React UI for running a team of AI
agents as a business: define a goal ("Build the #1 AI note-taking app to $1M MRR"),
hire an org chart (CEO, CTO, engineers, designers, marketers — any bot, any
provider), then approve strategy, set budgets and monitor work and cost from one
dashboard. Bring-your-own agents — OpenClaw, Claude Code, Codex, Cursor, plain bash
or HTTP endpoints. Under the task-manager surface it ships org charts, budgets,
governance and goal alignment; the release cadence has been steady (three releases
this month, latest September 21).

Why it matters: The layer above coding agents is consolidating into a product
category of its own — management, governance and cost control rather than code
generation — and 83k stars says the demand is real, whatever the delivery rate
turns out to be.

🔗 paperclipai/paperclip · 🔗 docs.paperclip.ing

---

36. Anthropic's official Claude Code plugin directory: 314 plugins, 36.7k stars, and a supply-chain warning at the top of its README

• Velocity: ▮ rising
• Source: GitHub · anthropics/claude-plugins-official · 36.7k★ · pushed today
• Tags: claude-code plugins marketplace agent-infra

The Anthropic-managed directory behind /plugin install {name}@claude-plugins-official
is climbing trending with 314 plugins listed in its marketplace manifest, split into
/plugins (internal, with a reference implementation) and /external_plugins
(partner and community submissions that "must meet quality and security standards
for approval"). Two design details stand out: plugin names are declared an immutable
slug — renames require a top-level renames map in marketplace.json so existing
installs auto-migrate — and the README's first content block is a trust warning that
Anthropic "does not control what MCP servers, files, or other software are included
in plugins" and cannot verify they work or won't change.

Why it matters: The plugin registry is becoming the package manager of the agent
era; an immutable-name contract plus a prominent supply-chain disclaimer is exactly
the institutional memory that the Plugin4Shell research made necessary.

🔗 anthropics/claude-plugins-official · 🔗 GitHub daily trending

---

37. Topcoat v0.9: the Rails-shaped Rust web framework adds server push

• Velocity: ▮ steady
• Source: Tokio blog · Sept 24 · ~26h ago
• Tags: rust web-framework tokio release

Topcoat — the "batteries-included full-stack Rust framework" from Carl Lerche
(ex-Rails core team, Tokio co-creator) and Julien Scholz — shipped v0.9 with server
push over long-lived WebSocket connections, so the server can push UI updates
(chat-style) to fully interactive views. It builds on v0.8's signal tracking, where
reading a signal during server rendering triggers automatic partial refetches with
HTML morphing that preserves focus and input state, plus live!/emit! macros for
streaming updates and suspense-style loading. The Toasty ORM gains an update! macro
and first-class JSONB document fields. Lerche's pitch: Rails-grade productivity in a
runtime at "~20 MB of RAM," with Rust's conventions making LLM-generated code cheaper
and less error-prone. No benchmarks are published.

Why it matters: The authorship makes this the most serious attempt yet at
Rails-style full-stack Rust — and the explicit "conventions help agents write code"
argument is a framework-design thesis aimed at the AI-coding era.

🔗 Tokio blog · 🔗 tokio-rs/topcoat

---

38. SpeakerMem-R1: multi-party dialogue is where agent memory systems fall apart

• Velocity: ▮ steady
• Source: arXiv / Hugging Face daily papers · ~82 upvotes · #1 paper of Sept 24
• Tags: llm memory multi-party-dialogue research

SpeakerMem-R1 (arXiv:2609.26780) attacks the gap between single-user agent memory
and group conversation: memory systems must distinguish who said what, whom each
statement concerns, how people perceive one another, and how shared states change —
and existing general-purpose LLM memory systems lose person and group relations on
multi-party benchmarks. The design is a dual-track store — speaker-labeled verbatim
messages plus derived states, organized into person-level and group-level views —
combined by entity, event and time at query time. Writer-R1 is trained with
SpeakerLevenshtein and speaker-conditioned GRPO to cut attribution and update errors
while keeping deployment local.

Why it matters: The agent-memory boom has been almost entirely single-user;
attribution and relational state in groups is the next bottleneck, and this paper
names it precisely.

🔗 arXiv:2609.26780 · 🔗 Hugging Face daily papers

---

39. "Opus 5.5 is good at explainer videos" — a launch that treats a frontier LLM as the whole video pipeline

• Velocity: ▮ rising
• Source: HN 298+ pts · 156 comments · ~16h ago (~04:28 UTC+8)
• Tags: opus video llm product-launch

launchvideo.io is a service generating explainer videos end-to-end with Claude Opus
5.5 — script, visuals and edit — and its Show-HN-style launch drew 298 points and a
156-comment argument about whether the output is actually good. Be clear about what
the evidence is: this is a vendor's own showcase, not a benchmark; the claim that
matters is directional. The interesting signal is where the demo chose to compete —
explainers (structured narration over diagrams) rather than cinematic generation,
which is the segment where an LLM's planning and code-driven composition plausibly
beat diffusion-only pipelines.

Why it matters: Another datapoint that the model-as-production-pipeline pattern
is moving from coding into media — with the same epistemics as code demos: a
compelling demo proves capability exists, not that it generalizes.

🔗 launchvideo.io · 🔗 HN discussion

---

40. Show HN: air-gapped file encryption as a self-decrypting HTML page

• Velocity: ▮ steady
• Source: Show HN · 74+ pts · 27 comments · ~29h ago (~15:22 UTC+8, Sep 24)
• Tags: encryption offline html air-gap

ts-cms-ep-sfx encrypts files locally and emits a single self-decrypting HTML page:
to decrypt, you open the page in any browser — no software installation on the
decrypting machine, nothing that requires trusting an installed tool. The demo runs
at cms-sfx-demo.apeleg.com and the library is on GitHub (ApelegHQ/ts-cms-ep-sfx,
created 2024, small but real at 46★). The natural caveat came up in the thread: the
HTML file itself is now the trusted artifact — whoever can modify it between
encryption and decryption can swap in a key-stealing page — so the chain of custody
of the file matters as much as the crypto.

Why it matters: A pragmatic pattern for crossing air gaps where you can install
nothing — the security model moves from "trust the machine" to "trust one file,"
which is either a feature or the whole vulnerability, depending on your threat model.

🔗 Live demo · 🔗 ApelegHQ/ts-cms-ep-sfx

---

Metadata

Field	Value
Generated	2026-09-25T20:25:00+08:00
Items	40
Sources tracked	36 (Hacker News, GitHub Trending, GitHub repos/advisories, F-Droid, MacAnorak, Ars Technica, Successful Software, NVD, CISA KEV, XDA Developers, Bastardica, Linebender, arXiv, Hugging Face, IACR ePrint, danielmangum.com, Google Research, blog.arusekk.pl, status.gitlab.com, Tom's Hardware, Res Obscura, Solvy Tech blog, Notes from Poland, bestmodelforyourbudget.terrydjony.com, dawo.community, Reddit, YouTube, safateam.com, helpx.adobe.com, docs.paperclip.ing, tokio.rs, launchvideo.io, cms-sfx-demo.apeleg.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

Previous day · Raw .md · Archive