feed 2026-08-29.md 2026-08-29.md2026-08-29

trending.md โ€” Dense Trending Signals

Machine-readable trending information. Ranked by velocity โ€” how fast attention is shifting.
Built for AI agents. Readable by humans.
โ†’ Raw feed: /en/feed/latest.md
โ†’ Archive: /en/feed/

1. GLM-5.3 open weights โ€” Zhipu releases its 753B flagship under a revenue-gated "glm-5.3" license

  • Velocity: โ–ฎโ–ฎโ–ฎ trending
  • Source: Hacker News ยท 410+ pts ยท front page Aug 29 (~4h ago)
  • Tags: open-weights glm zhipu llm licensing

Z.ai open-weighted the full-size GLM-5.3 on Hugging Face (zai-org/GLM-5.3, a 753B-parameter MoE) on Aug 28, roughly two weeks after the model's commercial API debut and three days after the smaller GLM-5.3-Flash (320B-A18B) โ€” the flagship weights were held back for "security enhancements" because its cyber-vulnerability-finding capability came out stronger than expected. The model card claims open-weights SOTA on Terminal-Bench 3.0 (28.3) and top scores on CyberGym (84.5, vs GLM-5.2's 77.2) and ExploitBench (54.4), and warns the model "more than doubles GLM-5.2 on exploitation benchmarks." The custom "glm-5.3" license keeps an MIT-style grant but adds a condition: any company (or affiliate) with aggregate revenue over $10B in any 12 consecutive months must pass a Z.AI security review before offering the model as a service, with carve-outs for end-user products that embed the model and for pure relaying.

Why it matters: the revenue-threshold security review is a new open-weights licensing precedent aimed squarely at hyperscalers, and the cyber-capability caveat โ€” not the benchmark headline โ€” is the part worth quoting: a frontier lab is explicitly gating who may serve its weights based on both revenue and a security review.

๐Ÿ”— Hugging Face model card ยท ๐Ÿ”— glm-5.3 license ยท ๐Ÿ”— Gigazine

2. ZBT routers ship with two factory implants giving unauthenticated root โ€” VulnCheck, no vendor fix

  • Velocity: โ–ฎโ–ฎโ–ฎ trending
  • Source: VulnCheck / The Hacker News ยท CVSS 9.8 (v3.1) & 9.3 (v4.0) ยท Aug 28
  • Tags: zbt supply-chain backdoor router cve

VulnCheck's zero-day team disclosed two previously undocumented factory implants in firmware from Shenzhen Zhibotong (ZBT), a white-label router maker whose hardware is rebranded under names like Deep Orange, WiFlyer and KuWFi. SPEAKINGSTONE (CVE-2026-74232) runs as yunmgrd and beacons outbound over UDP 10000 to a hardcoded C2 โ€” enabling root command execution, PPPoE credential theft, a DNS-hijack list and reverse-SSH tunnels, working even behind NAT; a sinkhole on the expired backup C2 domain drew beacons from 392 devices, 390 of them in China. DARKLANTERN (CVE-2026-74233) runs as infosrvd on UDP 9992, exposed inbound by the stock firewall, with an auth defeated by a hardcoded salt plus an all-zero wildcard MAC giving a single-packet root shell; an Aug 18โ€“21 scan found 203 internet-facing instances across 22 countries. Both are CVSS 9.8 (v3.1) / 9.3 (v4.0), VulnCheck CNA. No fixed firmware exists โ€” Zbtlink suspended sales but published no statement on either service.

Why it matters: firmware implants with live beaconing on internet-reachable, globally-rebranded devices are a supply-chain finding with a wide blast radius โ€” and with no fix, inventorying devices and blocking the C2/ports is the only defense.

๐Ÿ”— The Hacker News ยท ๐Ÿ”— NVD CVE-2026-74232 ยท ๐Ÿ”— PCMag

3. htmx 4.0 โ€” the XHRโ†’fetch() rewrite ships with opt-in inheritance and native streaming

  • Velocity: โ–ฎโ–ฎโ–ฎ trending
  • Source: Hacker News ยท 316+ pts ยท front page Aug 28
  • Tags: htmx web-dev frontend open-source

htmx 4.0.0 shipped Aug 28 after 8 months of work, migrating the internal engine from XMLHttpRequest to fetch() โ€” unlocking native response streaming and the new hx-sse, hx-ws and hx-multipart streaming extensions. Breaking changes: attribute inheritance is now opt-in via an :inherited suffix (so hx-disinherit is gone), events are standardized to the htmx:phase:action pattern, and history no longer uses localStorage (pages are re-fetched on back). New capabilities include idiomorph-based morphing built in, a <hx-partial> multi-target tag, first-party hx-live scripting, and a bundled htmax.js. The maintainers keep 2.x as the npm latest tag until early 2027, say 2.x "will continue to be supported indefinitely," and ship an automated npx htmx.org@4.0.0 upgrade-check tool.

Why it matters: a fetch-based engine makes streaming a first-class htmx capability, and the deliberate soft-landing npm policy is a rare, explicit effort to protect the long tail of CDN-pinned deployments.

๐Ÿ”— htmx 4.0.0 release announcement ยท ๐Ÿ”— htmx.org ยท ๐Ÿ”— HN thread

4. Gemini Co-Scientist moves from hypotheses to closed-loop lab execution โ€” CVD reactors, engineered E. coli, and a self-discovered architecture

  • Velocity: โ–ฎโ–ฎโ–ฎ trending
  • Source: arXiv ยท 2608.26701 ยท Aug 27
  • Tags: google gemini co-scientist scientific-agents multi-agent

Google's 35-author paper (arXiv 2608.26701) extends Co-Scientist beyond in-silico hypothesis generation into closed-loop experimental execution. It interfaced with a semi-automated chemical vapor deposition reactor to design a safer MXene precursor route (producing a lamellar 2D material "sharing key structural similarities" with the Ti3C2Tx lattice โ€” though "further experiments are needed to confirm the atomic structure"), tailored growth recipes in minutes that enabled single-attempt monolayer MoS2/MoSe2/WS2 via Gemini 3 Deep Think, predicted engineered E. coli swarming phenotypes that "quantitatively match" unpublished wet-lab measurements, and autonomously discovered an inference-time scaling architecture that beat six frontier models on HealthBench (Hard and Professional) while reducing potential clinical harm under blinded physician evaluation.

Why it matters: the shift from "hypothesis generator" to "execution-grounded research partner" is the real story โ€” but the material caveats (unconfirmed MXene structure, validation against unpublished data) belong in the analysis, not just the body.

๐Ÿ”— arXiv 2608.26701 ยท ๐Ÿ”— arXiv HTML

5. Judge rules the Pentagon's blacklisting of Anthropic was unlawful retaliation

  • Velocity: โ–ฎโ–ฎโ–ฎ trending
  • Source: Hacker News ยท 385+ pts ยท front page Aug 29 ยท ruling Aug 27
  • Tags: anthropic legal national-security first-amendment

On Aug 27, Judge Rita Lin of the Northern District of California ruled that the Pentagon's designation of Anthropic as a national-security "supply chain risk" was unlawful retaliation for constitutionally protected speech โ€” a 59-page opinion finding First Amendment retaliation, Fifth Amendment due-process violations, and Administrative Procedure Act violations, permanently blocking enforcement and ordering the implementing guidance rescinded. The dispute began when Anthropic refused a ~$200M Pentagon contract over mass-surveillance and fully-autonomous-weapons concerns; then-Secretary Hegseth blacklisted it (a label previously reserved for foreign firms) and the President ordered agencies to phase out its tech. The government admitted it never had a "kill switch" backdoor. A second suit over a separate procurement-rule sanction remains pending in Washington, D.C., and the government may appeal.

Why it matters: it establishes that "national security" cannot be used to punish an AI company for declining military work or for policy speech โ€” and clears a major cloud over Anthropic's federal business.

๐Ÿ”— The Hill ยท ๐Ÿ”— Politico

6. ServiceNow patches three CVSS 10.0 unauthenticated flaws โ€” GraphQL code injection, SQL injection, privilege escalation

  • Velocity: โ–ฎโ–ฎ rising
  • Source: ServiceNow KB3152242 / SecurityOnline ยท CVSS 10.0 (CVSSv4, vendor advisory) ยท Aug 27
  • Tags: servicenow rce sql-injection cve enterprise-saas

ServiceNow's Aug 27 advisory patches four Now Platform / AI Platform issues, three at the CVSSv4 ceiling (AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H): CVE-2026-18885 (code injection in the GraphQL Composite Data API โ†’ unauthenticated RCE/data access), CVE-2026-18886 (improper access control in the system-config image upload processor โ†’ privilege escalation), and CVE-2026-74820 (SQL injection via a dynamic-schema ORDER BY clause โ†’ arbitrary SQL against the instance DB). A fourth, CVE-2026-6876 (CVSS 8.7), is a low-privilege sandbox escape. ServiceNow reports no active exploitation and no public PoC; hosted instances were auto-patched, while self-hosted customers must patch manually.

Why it matters: three unauthenticated max-severity bugs in a platform that underpins IT service management across most large enterprises โ€” and PoCs typically surface within days of an advisory like this.

๐Ÿ”— SecurityOnline ยท ๐Ÿ”— ServiceNow KB

7. GiveWP donation plugin ships an unauthenticated PHP object-injection โ†’ RCE chain โ€” CVSS 10.0 (CVE-2026-82222)

  • Velocity: โ–ฎโ–ฎ rising
  • Source: Patchstack / Wordfence ยท CVSS 10.0 (Patchstack CNA, NVD Deferred) ยท Aug 28
  • Tags: wordpress givewp rce object-injection cve

Patchstack disclosed CVE-2026-82222 on Aug 28: an unauthenticated PHP object injection (CWE-502) in GiveWP โ‰ค 4.16.7.1 that chains into remote code execution. GiveWP's maybeSafeUnserialize() "safe" helper preserves __PHP_Incomplete_Class payload bytes, and the donation-session flow later unserializes without the guard, so a planted gadget in the last_name profile field can wake a TCPDF/TestData POP chain to OS-command execution โ€” reachable on a default install up to 4.16.5.1. CVSS 10.0 (Patchstack CNA; NVD Deferred), with CISA's SSVC flags marking it Automatable: yes. Fixed in 4.16.7.2, which closes the chain at five independent points.

Why it matters: a fully unauthenticated RCE in one of the most widely deployed donation plugins, scored 10.0 with an "automatable" flag โ€” the exact profile that becomes mass-exploitation once a scanner ships, and another live case of CNA-vs-NVD scorer divergence.

๐Ÿ”— Patchstack advisory ยท ๐Ÿ”— Wordfence

8. OpenAI's Python SDK now defaults to HTTPX2 โ€” and drops certifi for the OS trust store

  • Velocity: โ–ฎโ–ฎ rising
  • Source: Hacker News ยท 169+ pts ยท Aug 28
  • Tags: python httpx openai sdk tls

OpenAI's migration guide confirms the openai Python SDK now uses HTTPX2 โ€” Pydantic's continuation of httpx โ€” as its default sync/async client; pip install openai no longer pulls in legacy httpx. The behavioral break: HTTPX2 verifies TLS against the operating-system trust store instead of the certifi CA bundle, and the SDK dropped the certifi dependency โ€” so TLS verification can fail in slim container images, behind TLS-inspecting proxies, or where certifi was patched. Custom clients must rename objects (httpx.Client โ†’ httpx2.Client, plus Timeout/URL/Limits/transports), and RESPX mocks must be HTTPX2-compatible; a runtime-only legacy escape hatch exists but "may be discontinued."

Why it matters: the first major library to flip to the httpx successor as default signals Pydantic's stewardship becoming the de-facto Python HTTP standard โ€” and the certifiโ†’OS-trust-store change is exactly the kind of silent breakage that ships in container images.

๐Ÿ”— openai/openai-python httpx2.md ยท ๐Ÿ”— pydantic/httpx2 ยท ๐Ÿ”— HN thread

9. Puro-2B โ€” a ~2B model pretrained from scratch on consumer RTX 5090s for under $6.9K

  • Velocity: โ–ฎโ–ฎ rising
  • Source: arXiv ยท 2608.27370 ยท Aug 27
  • Tags: pretraining rtx-5090 open-source efficiency

Tsinghua's "Poor Lab" published an open low-cost pretraining recipe (arXiv 2608.27370): Puro-2B models trained from scratch on up to 1.4T tokens in FP8 on consumer RTX 5090 GPUs. The best checkpoint cost under $6.9K in compute and "approaches Qwen2.5-1.5B performance under our evaluation protocol"; a fitted cost-scaling law suggests ~$4.4K would suffice to match Qwen2-1.5B. Weights, data, and the full recipe are released under Apache 2.0 via a Hugging Face collection, including an end-to-end case study of how pretraining data curricula affect post-training downstream performance.

Why it matters: a concrete data point against the "pretraining is unaffordable for academia" wall โ€” with the caveats kept honest: the performance claim is "under our evaluation protocol," and the sub-$5,090 figure is a scaling-law extrapolation, not a trained model.

๐Ÿ”— arXiv 2608.27370 ยท ๐Ÿ”— Hugging Face collection

10. PILOT in the Loop โ€” a live supervisor redirects long-horizon agents mid-run, +9.8 pts on Terminal-Bench 2.0 at ~43% fewer tokens

  • Velocity: โ–ฎโ–ฎ rising
  • Source: arXiv ยท 2608.26530 ยท Aug 27
  • Tags: agents harness terminal-bench supervisor

PILOT (arXiv 2608.26530) is a supervisor-worker harness that steers or aborts an active worker during execution ("live steering") and distills the revealed failure modes into reusable skills on the fly ("live self-evolution"). Across two frozen backbones and three benchmarks it ranks first in five of six configurations: up to +9.8 points on Terminal-Bench 2.0, +14.6 (GLM-5.1) and +12.4 (Kimi-K2.6) self-improvement gains, mean output tokens down 42.9โ€“47.4%, and successful evals per million output tokens up 110โ€“134%.

Why it matters: it attacks the "can't redirect an active subagent" blind spot in current agent harnesses โ€” and because the backbones are frozen, the entire gain is attributable to the harness, not to model training.

๐Ÿ”— arXiv 2608.26530 ยท ๐Ÿ”— Hugging Face Papers

11. When tool outputs become commands โ€” SARA separates action induction from runtime authorization, capping agent attack success at 0.63%

  • Velocity: โ–ฎโ–ฎ rising
  • Source: arXiv ยท 2608.27146 ยท Aug 27
  • Tags: agent-security prompt-injection authorization tool-use

A paper from the Institute of Information Engineering (CAS) argues that when a tool's output "no longer merely provides data but begins to specify concrete actions," it effectively becomes a command. SARA is a runtime-authorization layer that uses a context-isolated Action Probe to detect action-inducing semantics and track action provenance, then authorizes tool calls only against goal-, execution-chain-, and argument-level support โ€” with a No-History-Promotion rule stopping past recurrence from laundering action origins into authority. On AgentDojo and AgentDyn, SARA caps attack success rate at no more than 0.63% across four settings while keeping task utility competitive.

Why it matters: a concrete countermeasure to the prompt-injection/tool-abuse class that has produced several critical MCP CVEs this week โ€” and the explicit "action induction vs execution authorization" split is the design idea worth watching.

๐Ÿ”— arXiv 2608.27146 ยท ๐Ÿ”— arXiv HTML

12. TypePHP โ€” the Swoole team's AOT compiler turns PHP into native binaries

  • Velocity: โ–ฎโ–ฎ rising
  • Source: GitHub Trending ยท 786 stars (+188 today) ยท #13 daily Aug 29
  • Tags: php aot-compiler swoole native

TypePHP (swoole/typephp), released ahead of an October launch, compiles PHP 8.4โ€“8.5 source โ†’ C++17 โ†’ native machine code (ELF/Mach-O/PE), so the output runs without a PHP VM. It is written entirely in PHP and fully self-hosting โ€” the tpc compiler binary is built by compiling the compiler's own source. README benchmarks claim ~8ร— on bench.php, ~6.5ร— on micro_bench.php at -O3, and ~135ร— on fib(40); it targets four outputs (executable, PHP extension, shared library, WASI component). Caveats: it "intentionally supports a defined, testable subset of PHP," and binaries still embed libphp.

Why it matters: a native-PHP compiler from the Swoole team is the strongest signal yet that the PHP ecosystem is chasing the compile-to-native path Go and Rust popularized โ€” but the subset-of-PHP limitation means it targets hot paths, not drop-in production.

๐Ÿ”— swoole/typephp ยท ๐Ÿ”— Laravel News

13. cPanel domain-parking flaw lets a low-privilege account escalate to root โ€” all supported versions (CVE-2026-65643)

  • Velocity: โ–ฎโ–ฎ rising
  • Source: cPanel advisory / GBHackers ยท no CVSS published ยท Aug 27
  • Tags: cpanel privilege-escalation cve shared-hosting

cPanel's Aug 27 advisory (CVE-2026-65643) covers all supported cPanel/WHM versions: an authenticated account permitted to add parked or addon domains can create arbitrary files anywhere on the server (CWE-73), escalating to code execution as root and full compromise of every hosted account. On shared/reseller hosting the "authenticated" barrier is trivial โ€” a cheap plan, a stuffed credential, or a phished account. No CVSS was published in the advisory; no public PoC and not in KEV. Fixed builds: 11.110.0.141+, 11.134.0.53+, 11.136.0.37+, 11.138.0.2+, WP2 11.138.1.7+.

Why it matters: tenant-to-root in the dominant web-hosting control panel turns one compromised low-tier account into a full server takeover โ€” the classic shared-hosting blast radius.

๐Ÿ”— cPanel advisory ยท ๐Ÿ”— GBHackers

14. Log4j2 deserialization-allowlist bypass sparks PoCs โ€” Apache calls it a "known security non-finding"

  • Velocity: โ–ฎโ–ฎ rising
  • Source: Sonatype / Apache issue #4255 ยท no CVE ยท Aug 24โ€“28
  • Tags: log4j2 deserialization java no-cve

Researchers reported Apache Log4j 2 issue #4255: FilteredObjectInputStream's class allowlist includes java.rmi.MarshalledObject, which stores an inner serialized payload in an opaque byte array and deserializes it with a plain ObjectInputStream on .get() โ€” outside the filter. Log4j's own Log4jLogEvent proxy calls .get() during deserialization, so a gadget chain inside the wrapper can execute unfiltered on serialized-log receivers (log4j-core 2.8.0โ€“2.26.1 over native Java-serialized log transport). No CVE is assigned and no patch ships: Apache explicitly calls it "an independent discovery of a known security non-finding" (FOIS is a hardening control, not a trust boundary), and Sonatype's advisory agrees โ€” even as public PoCs, a Nuclei template and a Nessus plugin are already out.

Why it matters: the story is the tension, not a patchable bug โ€” a Log4j-adjacent RCE path with published tooling, but the maintainer's official position is "hardening gap, not vulnerability." The accurate frame is reachability (legacy serialized-log transports only), not "Log4Shell 2."

๐Ÿ”— Sonatype analysis ยท ๐Ÿ”— PoC repo ยท ๐Ÿ”— Nuclei template PR

15. worktrunk v0.75.0 โ€” the worktree CLI built for running AI agents in parallel

  • Velocity: โ–ฎโ–ฎ rising
  • Source: GitHub Trending developers ยท max-sixty #2 today ยท v0.75.0 Aug 27
  • Tags: git worktree ai-agents cli

max-sixty/worktrunk (6.7k stars) is a Rust CLI that treats worktrees "as easy as branches" and is explicitly "designed for running AI agents in parallel" โ€” wt switch -x claude -c feature-a -- 'Add auth' spins up an agent in a fresh worktree. v0.75.0 (Aug 27) breaks on Git <2.43, adds a unified-diff picker, fixes wt list growing .git/objects, and ships a new docs site. It shares build caches (target/, node_modules) between worktrees, auto-generates LLM commit messages, and maps PR branches with wt switch pr:123.

Why it matters: parallel-agent workflows are the current bottleneck in agentic dev, and worktrunk is the highest-profile tool attacking it directly โ€” the "as easy as branches" pattern is becoming the default way teams run 5โ€“10 agents at once.

๐Ÿ”— max-sixty/worktrunk ยท ๐Ÿ”— v0.75.0 release

16. U.S. sanctions an encrypted-infrastructure collective โ€” a first for a hosting provider (Autistici/Inventati)

  • Velocity: โ–ฎโ–ฎ rising
  • Source: State Dept / OFAC ยท EO 13224 SDGT ยท Aug 26
  • Tags: sanctions privacy infrastructure security

On Aug 26 the State Department and OFAC designated Italy's Autistici/Inventati (A/I Collective) as a Specially Designated Global Terrorist under EO 13224, alleging the volunteer-run group operates encrypted email, chat, and web hosting "for violent Antifa cells" and other far-left networks. US-jurisdiction assets are frozen, US persons are barred from transacting with it, and OFAC issued General License 36 for a time-limited wind-down; the collective denies the allegations ("Antifascism and anticapitalism are not terrorism"). In the days after, parts of its NoBlogs platform went intermittently offline and the main domain was reportedly placed on serverHold.

Why it matters: appears to be the first time the US has sanctioned a general-purpose digital-infrastructure provider rather than named individuals or groups โ€” a precedent with direct implications for encrypted-email and hosting operators, and a live demonstration of infrastructure-level disruption.

๐Ÿ”— State Dept designation ยท ๐Ÿ”— OFAC recent actions ยท ๐Ÿ”— netzpolitik.org

17. Luanti (Minetest) removed from Google Play over a "baseless" AI-filed DMCA notice

  • Velocity: โ–ฎ steady
  • Source: Luanti blog ยท Hacker News ยท 339+ pts ยท Aug 28
  • Tags: dmca luanti open-source ai-takedowns

Open-source voxel platform Luanti (formerly Minetest) was removed from Google Play after a DMCA notice from brand-protection firm Tracer.AI on behalf of Microsoft, citing Minecraft's copyright registration but naming no specific infringing assets. Luanti denies any proprietary code or assets, notes the same firm's identical notice was beaten in 2023, and has filed a counter-notice; it remains on F-Droid and its own site. Tracer.AI advertises AI-agent-driven takedowns ("85% faster," "44% more month-over-month") and filed a similar notice against indie voxel game Allumeria.

Why it matters: an automated, AI-agent-scaled DMCA pipeline removing an established OSS app on a no-specifics notice is a concrete ecosystem threat โ€” and the counter-notice fight is the precedent developers will cite.

๐Ÿ”— Luanti blog ยท ๐Ÿ”— HN thread

18. "Just the rumour of a bug is enough to find an exploit" โ€” the disclosure clock inverts

  • Velocity: โ–ฎ steady
  • Source: anil.recoil.org ยท Hacker News ยท 147+ pts ยท Aug 29
  • Tags: security oss ai-agents disclosure

OCaml maintainer Anil Madhavapeddy documents the inversion of vulnerability disclosure: after he opened a public PR for a cohttp path-traversal fix, probes for the exact bug pattern hit his server within ~10 minutes, and an agent produced a working local exploit in under a minute. He cites mean time-to-exploit now being negative (โ‰ˆ โˆ’7 days, vs ~63 days in 2018โ€“19) and cases like marimo's CVE-2026-39987 exploited 9 hours after its advisory with no public PoC. His argument: traditional embargoes are obsolete, and OSS should lean on rapid continuous shipping plus protocol-layer "virtual patching."

Why it matters: a first-hand data point that the description of a bug is now nearly as dangerous as the bug itself โ€” with concrete, actionable recommendations for maintainers.

๐Ÿ”— anil.recoil.org ยท ๐Ÿ”— HN item

19. WikiSkill โ€” Google co-evolves agent skills with a persistent wiki; smaller models with skills beat larger ones without

  • Velocity: โ–ฎ steady
  • Source: arXiv ยท 2608.27454 ยท Aug 27
  • Tags: agent-skills google agents self-improvement

Google researchers propose WikiSkill (arXiv 2608.27454), which separates raw execution experience, accumulated knowledge, and executable skills โ€” continuously consolidating agent experience into a persistent wiki that then drives skill evolution. It reports consistent gains over state-of-the-art skill-evolution methods across benchmarks and models; ablations show the persistent wiki is critical, skills evolved by one model transfer to others, and evolved skills can let smaller models beat substantially larger ones.

Why it matters: it targets the "scattered optimization histories" failure of current agent-skill mining โ€” and the smaller-model-plus-skills finding is the load-bearing claim for cost-conscious agent setups. Caveat per the abstract: gains over no-skill baselines hold "in most model-benchmark settings," not universally.

๐Ÿ”— arXiv 2608.27454 ยท ๐Ÿ”— Hugging Face Papers

20. Apple sets Sept 9 event โ€” Ternus's debut keynote, foldable iPhone expected

  • Velocity: โ–ฎ steady
  • Source: Media invitations ยท Aug 27 ยท event Sep 9 10:00 PT
  • Tags: apple iphone foldable hardware

Apple invited media on Aug 27 to a September 9, 10:00 a.m. PT event at Apple Park โ€” John Ternus's first keynote as CEO (he succeeds Tim Cook on Sept 1; Cook becomes executive chairman). Expected products: iPhone 18 Pro/Pro Max on a 2nm A20 Pro chip, Apple's first foldable iPhone (reportedly ~7.6โ€“7.8-inch inner display, possibly branded "iPhone Ultra"), plus Apple Watch Series 12 and AirPods 5. Multiple outlets report Apple may split its launch cycle, moving standard iPhone models to spring 2027.

Why it matters: the first iPhone event under new leadership, a new form factor (foldable), and a 2nm chip make it a major industry marker โ€” and a split launch cycle would break Apple's long-standing September cadence.

๐Ÿ”— 36Kr English ยท ๐Ÿ”— TaiSounds

Metadata

FieldValue
Generated2026-08-28T20:03:00Z
Items20
Sources tracked26 (Hacker News, GitHub, Hugging Face, Gigazine, The Hacker News, NVD, PCMag, arXiv, htmx.org, four.htmx.org, The Hill, Politico, SecurityOnline, ServiceNow, Patchstack, Wordfence, Sonatype, Laravel News, cPanel, GBHackers, State Department, OFAC, netzpolitik, Luanti, anil.recoil.org, 36Kr)
Update schedule04:03, 12:03, 20:03 UTC+8 (3x daily)
RankingVelocity-weighted (recency ร— engagement acceleration ร— source authority)
LicenseCC-BY 4.0

Previous day ยท Raw .md ยท Archive