Platform gatekeeping — capability removal and distribution chokepoints (Sep 2026)

A pattern running through the Sep 2026 batches: platform owners resolving abuse (malicious extensions,
fraudulent installs, ad friction) by removing capability classes — taking legitimate tools down with
them — and by gating distribution for clients that depend on the platform's own endpoints. The parties
hit hardest are precisely the anonymous, accountless, ad-blocking, privacy-first users the platform doesn't
monetize.

Chrome removes the last Manifest V2 extensions — user-agent blocking ends in Chromium

• Google "reached the final milestone": all remaining MV2 extensions have been removed from the Chrome Web Store, and on Chrome ≤138 installed MV2 extensions keep running but can no longer be updated or reinstalled.
• The blast radius extends past Chrome: Brave and other Chromium forks rely on the CWS for discovery and install. Brave is now self-hosting four MV2 extensions (AdGuard, uBlock Origin, uMatrix, NoScript) on its own backend — the maintenance bill every Chromium fork now faces is self-hosted MV2 distribution or a move to declarativeNetRequest (weaker, allowlist-driven blocking).
• Caveats: the write-up is a small independent blog, not Google's own announcement; the HN thread (737 pts / 575 comments) is mostly "switch to Firefox" resignation — but nobody in it disputes that the removal happened.
• The timing is the framing: days after the "Superior" trojanized-extension campaign (Aug 30), the ecosystem's answer to malicious extensions was to remove the capability class — taking the legitimate blockers with it.

Firefox for iOS ships a built-in ad blocker — accommodation via the sanctioned API

• The counter-move: Mozilla added an optional, off-by-default Ad Blocker to Firefox for iOS built on Apple's WebKit Content Blocker API + EasyList — no extension system required, which is the point, since iOS doesn't support Firefox's desktop/Android extension model. Mozilla is explicit about limits: first-party ads, search ads, and Firefox's own sponsored new-tab shortcuts still appear.
• The rollout stumble is the case study: progressive rollout where enabling initially required "remote improvements" (a telemetry flag) — one commenter distilled it as "you can only block ads if you allow Mozilla telemetry" (Firefox 148 has since split remote improvements from telemetry), and many users still don't see the toggle. Ship the feature off-by-default, then require telemetry to find the switch.
• Significance: ad blocking inside a mainstream non-Safari iOS browser, done entirely with WebKit's public content-blocker API — exactly what Apple's platform constraints permit, which is precisely why it's possible.

Play Store blocks Aurora Store — anonymous installs die by credential-pool flagging

• Aurora Store (FOSS Play Store client) returns "Server busy, please try again later" on all installs via anonymous accounts — verified in the project's GitLab issue #1566 (filed Aug 31, Fairphone 5 / CalyxOS nightly), persisting across VPN changes, cache clears and account refreshes.
• The cause is unconfirmed: the leading explanation (a top HN commenter's) is that Aurora pools burner Google accounts for anonymous downloads and Google has flagged them; Google has made no statement. Framing correction applied per the Void lesson: GrapheneOS actually recommends the sandboxed Play Store, not Aurora — the real victims are account-free Android setups (CalyxOS, Sailfish-style) and anyone deliberately avoiding a Google account.
• The structural fact: a platform can de-facto kill account-free app installation by flagging one client's shared credential pool — with no appeal path for a project it doesn't host and no stated policy to comply with.

The pattern

1. Abuse becomes the justification for capability removal — trojanized extensions → MV2 gone; burner accounts → anonymous installs gone.
2. The legitimate users hit hardest are the unmonetized ones — ad blockers, accountless phones.
3. The surviving paths are sanctioned-API accommodation (WebKit Content Blocker) or self-hosted distribution (Brave's MV2 mirror) — both cost more than what was removed.
4. This is the client-side twin of the AI-crawler tax (open-infra-crawlers): anonymous, unauthenticated, non-monetized access is the platform's least-favored access class, and it is degrading everywhere at once.

"Hang on to Your Firefox" — retention sentiment becomes a measurable force (09-02)

• Mark Rogers (newsonaut.com): Firefox is "our last best hope for browser engine diversity and competition"; its shrinking share is why it deserves support rather than pile-on criticism; and the alternatives critics name (Vivaldi included) share the sins they cite. 722 HN points in eight hours — a mood reading, arriving the day after Chrome's MV2 removal (which the essay never mentions; its case is the engine itself, not extensions).
• Kept separate from the MV2 story on purpose: the argument predates and outlasts that trigger — conflating them is how aggregate framing errors start. The essay has visible seams (hedged conjecture on why Firefox joined X; a self-undercut Google-bots speculation), but the audience for "last independent engine" arguments just got much larger, and retention sentiment around Mozilla is now something a feed can measure as a force alongside the capability-removal events in this file.

Weedout — post-MV2, user-side curation lives on platform-native surfaces (09-02)

• A $1.99 Safari extension (masteranza.github.io, HN 157) removes YouTube videos carrying the platform's own "Made with AI" label from feed, search, related videos, playlists and Shorts, with optional Shorts auto-skip and a "Dim mode" that fades flagged items in place for verification before removal. Detection is deliberately non-clever: it filters only on YouTube's own disclosure badge — "no guessing, no heuristics, no false accusations" — processing locally in ~0.5s per live feed, no accounts, no data collection, one-time purchase.
• The stated limit is the entire product thesis: AI-made but unlabeled content is "out of scope (for now)." This fits the file's pattern from the consumer side: after Chrome's MV2 removal killed uBlock-class blocking, platform-native filtering surfaces (Safari content blockers, YouTube's own labels) are where user-side curation still lives — and trusting the platform's label is the price of admission. The HN thread runs the adjacent debate: whether hiding (rather than down-ranking) AI content changes what YouTube learns about you.

Registry-level namespace removal + the ToS gate on agent OAuth reuse (09-04)

• ICANN approved killing .name's third level — 22,000 personal domains disappear in February 2027. Verisign proposed on 2026-04-15 eliminating every third-level domain under .name — the only kind of domain .name has ever sold — and ICANN approved on 2026-07-28. Neil Fraser's writeup (970 HN points, the day's biggest story) is the first-person case: he loses fraser.name, an email address anchored to it for ~25 years, a website paid through 2040, and working IoT devices — and once the second level frees up, whoever registers it can impersonate him and reset every account tied to the old address. The gatekeeping shape at the registry layer: a namespace decision taken in July while almost nobody noticed converts long-lived personal identity roots into squatting targets. Sibling of security's dangling-delegation shape (the €5 ENUM domain): an identity anchored to something someone else controls carries that thing's expiry date.
• Google's Antigravity terms name OpenClaw as a bannable example. The Additional Terms of Service: "Using third party software, tools, or services to access the Service" is "a breach of this Agreement" — with the example text literally reading "using OpenClaw with Antigravity OAuth" — and "may be grounds for suspension or termination of your Antigravity and/or Gemini CLI accounts." Gergely Orosz's thread hit the HN front page, with community reports of suspensions for pointing agent harnesses at Antigravity/Gemini OAuth. A new leg of the pattern — not capability removal (MV2), not credential-pool flagging (Aurora Store), but contract-level gatekeeping: reusing a consumer AI subscription's OAuth in the harness of your choice is now a breach whose blast radius is the whole Google account. Direct collision with the BYO-subscription arbitrage trend (smart-routing's Sub2API / free-claude-code): the subscriptions being arbitraged have terms, and providers are starting to enforce them.
• Gmail removes "Send as" for third-party addresses (January 2027). Google's support page states it plainly: "Starting January 2027, Gmail will no longer support the 'Send as' feature for third-party email addresses, such as @yahoo.com or @outlook.com." Workspace aliases and other owned Gmail addresses are unaffected; no reason is stated anywhere on the page; the suggested alternatives are plus-addressing and Google Groups delegation. The 182-point HN thread is dominated by small businesses and people who run custom-domain mail through Gmail via external SMTP — for them the feature is the product (one top comment is a Workspace cancellation announcement). Another quiet consolidation of email identity into provider silos: authenticated send-through-arbitrary-SMTP inside a mainstream client dies with no migration path, deadline mid-Q1. The pattern's signature move again — the legitimate, unmonetized user takes the loss (MV2/Aurora Store/.name siblings).

The takedown loses to demand — Nitter regrows (09-05 20:03)

• A community-maintained Codeberg wiki now lists more working Nitter instances than existed before the takedown waves — the ecosystem rebuilt on a fork ("shitter"), some instances powered by bulk-purchased X accounts and residential proxies; XCancel's website is down but its RSS feeds still respond — the visible site was killed, the pipe wasn't. The thread's own caveats are part of the story: the wiki recommends an account gray-market seller one defender calls "extremely sketchy", and every instance listed is whack-a-mole-ephemeral — several commenters warn against linking to any of them durably (redirect tools, LibRedirect, or self-hosting behind basic auth are the stable paths).
• The gatekeeping pattern's inversion: suppression suppressed the instances but not the demand — account-free X reading is now a distributed gray-market arms race that regrows faster than it can be cut down. For anyone citing these links: they rot; treat the fork and the technique as the story, not any single instance.

The state acts on the provider itself — Autistici/Inventati shuts down (09-07 12:03)

• A/I — the Italian collective running privacy email/hosting (autistici.org / inventati.org) since 2001 — announced Sep 6 it is shutting down all services, with the OFAC General License 36 wind-down ending Sep 25. The endgame of the 08-29 designation: the US State Department designated A/I a Specially Designated Global Terrorist on Aug 26, 2026. The collective says its domain was made unreachable without notice and that continuing to operate would endanger its users and people near the collective; its announcement rejects the US government's characterization outright. Treasury and State press releases confirm the designation; the domain-seizure claim rests on A/I's own account, and the allegations are contested and unadjudicated. The HN thread (393 pts / 265 comments) runs on solidarity and on what US sanctions power means for internet infrastructure.
• Why it extends the pattern: every prior entry in this file was a platform removing a capability class or tightening a contract. This is the state acting directly on the infrastructure provider — the first counterterrorism designation aimed at a digital-infrastructure provider rather than a violent group — with direct spillover for any US-linked registrar, host or payment processor serving controversial communities: the compliance departments of those vendors are now the enforcement layer, no takedown notice required.

Nitter and XCancel resume service — 12 days after X Corp's cease-and-desist (09-07)

• The two largest X/Twitter frontends came back online Sep 6 (HN #3, 779 pts), twelve days after going dark Aug 24 on C&D letters that alleged "unlawful use and circumvention of X's API," cited the Texas Harmful Access by Computer Act and the Lanham Act, and demanded permanent takedown of all Nitter instances plus the code repository by Aug 25 5pm EST. Creator Zedeus took the services offline and stopped development "for the time being" while seeking legal advice, saying only they "won't be commenting further on the specifics" — then resumed without publishing any terms. The repo stayed up throughout.
• Why it matters: the first test of whether a C&D without litigation can permanently kill widely-used open infrastructure — the resumption implies counsel found the claims contestable enough to risk operating. Caveats are load-bearing: no public legal filing exists, any conditions of the resumption are undisclosed, and the whole account rests on Zedeus's statements. The HN thread reads the episode as a live probe of scraper liability under state computer-misuse laws. This closes the loop with the 09-05 note above: the takedown lost to demand once already (the "shitter" fork regrew); now even the original targets judged the letters survivable.

2026-09-10 — opaque automated flagging at the ad-platform chokepoint

• Google Ads flags a signed, notarized macOS app as "malicious software" (xlii.space Sep 9; HN 306+ pts / 185 comments). The developer of RACE (a signed-and-notarized native macOS terminal multiplexer) spent $500 on a first Google Ads campaign; the account was suspended for "Malicious software" and "Compromised Site" with zero detail on what triggered either flag. He documents exhaustive clean checks (Safe Browsing, Search Console, VirusTotal with hash, notarization, JS bundle review), four auto-rejected appeals, and a week-long block. A top-of-post edit after the HN thread: reinstated "through the apparent magic of Hacker News" — still no explanation. The gatekeeping shape, ad-platform edition: opaque automated security-flagging where the only recourse that worked was viral visibility, and reinstatement correlated with attention, not any evidence change. The author's own caveat holds: Safe Browsing clearance "does not establish what Google Ads detected."
• Sources: xlii.space writeup · HN discussion

2026-09-11 04:03 — "purchase ≠ ownership" gets its evidentiary genre

• Garcia v. Sony Interactive Entertainment (N.D. Cal., filed Jun 18, 2026): four plaintiffs allege the PlayStation Store's "Buy Now" framing violates California AB 2426, which bars implying unrestricted ownership of digital goods without clear license disclosure. A Consumer Rights Wiki page cataloguing Sony's own ownership language ("games you own," "verified owner") now enters as evidence — a crowd-maintained archive of a company's marketing copy as a litigation artifact, a new evidentiary genre for digital-ownership cases. Sony moved Aug 21 to compel arbitration (30-day ToS opt-out — no plaintiff opted out) or dismiss, arguing "reasonable consumers would not be misled"; hearing Oct 1 before Judge Vince Chhabria. Claims are allegations; Sony hasn't filed its merits reply, and the wiki doesn't claim Sony has removed the language — the HN headline slightly overstates. Joins this file as its ownership leg: the same platforms that remove capability classes also sell an impression of ownership the law now tests.
• Sources: Consumer Rights Wiki case page · HN discussion

2026-09-17 04:03 — the review queue itself becomes the bottleneck

• Google Play review now routinely takes longer than a week (HN 309+ pts, 295 comments): Daniel Gultsch (Conversations XMPP) documents review waits exceeding a week — Signal reports "4 hours to 5 days," CoMaps waited ~16 days — and attributes the backlog to AI-generated app spam flooding the review pipeline. The HN thread filled with corroborating timelines from other maintainers, including security releases stuck behind the queue and coordinated releases desynchronized. The gatekeeping angle: a review pipeline measured in weeks is a security-relevant bottleneck — it delays CVE fixes for millions of installed apps — and the suspected cause is the same generative wave filling the feed with skills filling the store with submissions. Google publishes no queue statistics; the evidence is maintainer testimony, broad but anecdotal. Mastodon permalink verified via the status API (created 2026-09-16T11:17:57Z).
• Sources: Daniel Gultsch on Mastodon · HN discussion

2026-09-22 20:03 — consent as a per-version state: "off" did not survive the upgrade

David Bushell's upgrade-day account (303-pt HN, 225 comments): he had explicitly disabled Apple Intelligence and Siri ("a rather explicit 'no'") after macOS 15.3 auto-enabled a feature phoning home every 15 minutes; upgrading macOS 15 → 27, he found the AI features re-enabled and the opt-out switch gone. "Disabled" Siri still leaves multiple unkillable Siri processes consuming memory and writing data; Apple Intelligence took 22.28 GB of disk (priced at ~£11 against Apple's £500/TB upgrades); the Screen Time hidden workarounds only hide AI features from menus without disabling them. This is the concrete answer to the question the macOS 27 item (09-22 12:03) left open — whether "off" means off: a prior opt-out evidently didn't survive the upgrade, so consent here is a per-version state, not a setting. Caveats carried: a single-user anecdote and an opinion essay — but every specific (processes, storage, Screen Time path) is checkable on the machine in question.

Sources: dbushell.com · HN discussion

2026-09-25 20:36 — ads on a paid OS with no off switch; moderation policy applied to product criticism

Apple adds persistent ads to iOS with no off switch (TechRadar; HN 122+ comments) — an ads-on-a-paid-OS escalation whose developer-side impacts (attention economics, referral flows) now run through Apple's own surface. Meta removed the Meta-glasses satire video (Roel Maalderink × Bits of Freedom — filmed with Meta's own glasses in Meta's Amsterdam office, staff discomfort being the point) from Facebook and Instagram under its bullying-and-harassment policy: the moderation rule built to protect users applied to suppress the argument that recording indicators don't make surveillance comfortable — product criticism enforced against via content policy; the YouTube version now carries the deletion in its title. GrapheneOS reports a "high chance" of phones shipping with it preinstalled in 2027 (Mastodon) — OEM distribution for the de-Googled hardened Android would be the counterweight datapoint to this whole ledger. And F-Droid 2.0's unified installer exists partly because the EU DMA made Android's pre-approval API available to non-commercial stores — regulation quietly working for the volunteer FOSS store (→ dev-tools).

Sources: HN — iOS ads · GrapheneOS on Mastodon · HN — Meta glasses video · F-Droid announcement

2026-09-26 20:03 — the paid-Play era ends for the canonical Android FOSS client; Cambridge Analytica gets a verdict

Conversations goes free (gultsch.de; HN 97 pts): Daniel Gultsch ends the canonical open-source Android XMPP client's paid Google Play distribution and makes the app free. His stated reasons: the 15% commission, support that goes nowhere ("there is no way to talk to a human at Google"), and no longer depending on the revenue — "For years I've felt like I was in a toxic relationship with Google, and the only reason I stayed was economic dependency… Google doesn't deserve me and my money anymore. I'm done." The app paid his rent for years; the HN thread split between "the commission funds app review" and "a monopoly extracting rent without accountability." The same author who documented the >1-week Play review queue on 09-17 now exits the store entirely — one of the few Android FOSS apps that ever made paid-Play distribution work is publicly abandoning the model, and the remaining monetization paths (donations, consortium funding) are both harder.

Facebook found liable in the Cambridge Analytica trial (New Mexico state jury, Santa Fe; verdict Sep 25; HN 261 pts): the jury found Facebook liable for deceiving users about privacy protections in AG Raúl Torrez's two-week trial centered on the personality quiz that harvested ~87M profiles — more than 2 million violations (New Mexico's whole population), including misleading the public about post-scandal data-broker investigations. The judge, not the jury, sets penalties; the state seeks the maximum $5,000 per violation. Context that narrows the field: August's up-to-$18B multistate child-safety settlement released Meta from future Cambridge Analytica liability everywhere except New Mexico (Florida declined to sign), and the state already won $942M in a separate minors-safety trial this year. Meta says it disagrees and will keep defending, arguing a First Amendment right to run its platforms as it sees fit. The engineering-relevant part is the per-violation penalty math: $5,000 × every affected user.

Sources: gultsch.de · HN — Conversations · CBS News · HN — verdict

2026-10-03 05:03 — the agent permission wall arrives on macOS; a court rejects a technical mandate on systems grounds

Apple will tighten Full Disk Access — and cites AI agents as the reason (developer notice Oct 2, no technical details): FDA "largely sidesteps" per-resource privacy controls and is being used "in ways that could put users at risk, exposing everything on their systems — including files, mail, messages, and even browsing history"; going forward, such access requires "very explicit user action." The AI framing is the headline: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." What the announcement does not contain: no effective date, no new entitlements or APIs, no migration guidance — "going forward" is the entire timeline. It also flags the third-party angle: agents reading communication apps "can also compromise the privacy of the people users are communicating with." The agent-era permission wall is arriving on macOS first, from the vendor that already gates everything else; if your agent indexes mail, messages or the filesystem, expect a consent cliff in a future macOS release — design for scoped access now, because "backup app" is the only use case Apple called legitimate.

Utah's VPN age-verification law blocked as "a technical impossibility" (EFF; 303 pts, #1): a federal court (Judge Barlow) granted a preliminary injunction against SB 73, which would have forced sites to block all VPN users or pierce traffic masking, with "commercially reasonable geolocation obfuscation detection" required from Oct 8. The holding is the rare court opinion written as a systems argument: the statute "requires entities like Aylo to geolocate its website users with perfection to avoid liability," while acknowledging "that geolocation perfection is not presently possible" — effectively strict liability for any single mis-located visitor. Brought by Aylo (Pornhub's parent) with EFF's amicus framing the technical record; scope limits stated: the injunction covers the VPN provisions only, a separate ban on sharing VPN circumvention information is unchallenged, and Utah may redraft next session. The first age-verification regime to die on a technical-impossibility holding rather than a speech ruling — the counter-case to this whole ledger: sometimes the court adopts the engineers' argument verbatim.

Sources: Apple Developer News · HN · EFF Deeplinks · HN

2026-10-04 04:03 — the state-side gatekeeping ledger gets document-level evidence: protest observers into a contractor-built system

ICE/Palantir ICM facial-recognition filing (Hilton v. Noem, D. Me., 2:26-cv-00092; partially unsealed filing public Oct 2): a DHS agent created Investigative Case Management records on at least six people (the government says eight) who observed ICE operations in Portland, Maine — labeling two "Threat to Law Enforcement, Professional Protestor" — and sent their photos to a CBP officer for facial-recognition checks via the Mobile Query app, shared as "lookout records" per a 2016 DHS privacy assessment. ICM is Palantir-built (2014, Gotham-based; five-year support contract up to ~$96M by 2022, plus $30M for ImmigrationOS in 2025). DHS's spokesperson: "the underlying lawsuit is based on the lie that there is a database." Status discipline: these are allegations in ongoing litigation, built heavily on the government's own documents and depositions; nothing is adjudicated, and the government's motion to dismiss states the agent "did not attempt to nominate any individuals to the terrorist watchlist." The story's real subject is the word being fought over: distributed case management with lookout-sharing behaves like a database without being called one — the gatekeeping ledger (SDGT on A/I, the Utah injunction) now includes surveillance infrastructure documented at the filing level.

Sources: Wired · CourtListener docket · HN discussion