F-Droid 2.0, Rails World's AI Pivot, and UK Two-Tier Encryption

Good morning. It's Friday, September 25, and we're covering the day the open web and the agent economy each moved: F-Droid 2.0 landed as the largest update in the store's 10-year history, and DHH used the Rails World keynote to argue that AI agents can now replace the very thing he built. Between them sit a UK encryption story with a two-tier privacy outcome, a GitHub support failure that only a Hacker News front page could fix, and the strongest sign yet that a frontier model can pitch, write, and render a video on its own.

The F-Droid 2.0 thread sits at 1,255 points with 362 comments, the UK two-tier encryption piece at 456 with 406, and the Rails World keynote at 338 with 368. Here is the signal, sourced.

In today's brief:

• F-Droid 2.0 is a full Kotlin Compose rewrite with a Discover/Search/My Apps navigation, and a unified installer built on Android's new pre-approval API - the DMA-era change that finally lets third-party stores install apps without the clunky consent loop
• DHH's Rails World keynote contained no Rails framework announcements; instead it pitched agentic coding with a report that Hey's backend was rewritten in Rust by AI agents and servers dropped from 110 to 10
• The UK's two-tier encryption is now permanent for Apple users: anyone who missed the window to enable Advanced Data Protection can never turn it on, while the tribunal fight over the secret order presses on
• Opus 5.5 is good at explainer videos: paste a URL and a serverless agent writes the film as code, then renders it deterministically - no video model involved, about 100k tokens per run

THE BIG ONE

F-Droid 2.0 Rewrites a Decade-Old Store to Survive the DMA Era

F-Droid announced what it calls its largest app update in 10 years: a complete redesign of the official client, rewritten in Kotlin with Jetpack Compose, rolling out to users over the coming weeks after 14 test releases. The navigation collapses to three surfaces - Discover, Search, and My Apps - with categories pulled into Discover, hundreds of finer-grained categories replacing the old broad buckets (Games alone splits into 17 genres), and search that now indexes descriptions, categories, and translated content with much stronger CJK support. It is the first big statement of how the free-software Android lane intends to stay alive now that the platform has shifted around it.

The install story is the structural change. F-Droid says that, "thanks largely to pressure from the EU's Digital Markets Act and anti-trust actions around the world," Android now offers third-party stores a smoother install path, and 2.0 uses it: a unified installer built on the new pre-approval API, letting the user confirm the install right after deciding, instead of after the download completes. That removes the biggest usability wedge between F-Droid and the built-in store. The trade is that F-Droid's Privileged Extension is not carried over, because the session installer can run background updates on recent Android without it, and the project dropped Android 6 after an upstream fix required Android 7.

Privacy features got more honest. Tor support moves to generic Proxy Settings with Tor VPN the recommended path, and the panic/app-hiding feature now only disguises the icon and name - F-Droid is explicit that the app still shows in Apps settings and is detectable under forensic inspection. One feature did not survive: the ability to wipe apps via a panic trigger, which the team says it will revisit if users who rely on it push back. The rewrite was funded by NLnet through Mobifree, the Open Technology Fund, NGI, and the Calyx Institute, and underwent an independent security review from the OTF Security Lab. It lands in the same Android-regulatory current as our recent on-device ADB coverage and EU interoperability analysis.

Why it matters: the competitive terrain for Android's everyman development channel just moved - a next-generation client plus a real install pipeline means F-Droid stops being the store you side-load once and forget, and starts being a distribution mechanism apps can actually ship through.

ENGINEERING

DHH's Rails World Keynote Was an Agentic-Coding Manifesto, Not a Rails Roadmap

The Rails World 2026 opening keynote (338 points, 368 comments) produced no Rails framework announcements at all - which is itself the story. Attendees' summaries paint it as DHH's case for agentic coding and native apps: Hey is moving to native desktop and mobile apps, and its backend was rewritten in Rust by AI agents, cutting the server fleet from 110 machines to 10. The most-quoted line is about the binary: "half a megabyte," with DHH observing it could almost fit on a floppy disk. The record this week gets complicated - some attendees disputed a secondhand claim that Hotwire was declared dead - but the direction is unambiguous, and DHH reportedly described himself as retired from being a professional programmer.

The HN reaction split roughly two ways. One camp read the talk as "the funeral of Ruby on Rails," including the highest-visible comments ("Comes to Rails conference. Gives a keynote with nothing about Rails. Leaves"). The other pushed back from inside the room: a front-row attendee called the vibe far from doom and gloom and noted the framework's own agentic-coding benchmarks page now ships on the official site, and the follow-up essay What About Rails? argues the framework is more relevant than ever precisely because agents are good at the boring parts. The pushing-Ruby-to-Rust move is worth framing soberly too: Hey moved rendering off the servers to clients, so the dramatic server reduction is not a like-for-like benchmark, and the pgrust story shows full Ruby-equivalent rewrites are a very different animal - a point we made in our pgrust deep dive and our Bun rewrite analysis.

Why it matters: the founder of the framework that defined a generation of web development is publicly betting his next products on agents writing the replacement code - the debate is no longer whether agents can write at scale, it is whether a language you expect them to eventually rewrite still deserves a community's identity.

SECURITY

GitHub Ignored a Malware-Toting Imitation of His App for Three Weeks - Until HN

Andy Brice of Successful Software (257 points, 112 comments) reported an imitation of his Windows and Mac app Easy Data Transform on GitHub on August 31: same product name, same logo, unlicensed. The repository's Mac .dmg returned malware warnings on VirusTotal and, in a flourish, the installer's background image had been swapped to one that told downloaders to ignore any warnings the OS showed. He filed the report the same day, got the automated acknowledgment, added the VirusTotal and image evidence on September 10, and heard nothing back for 23 days. The listing was taken down about ten minutes after the post hit the HN front page. "Total coincidence. I'm sure!"

The comment thread shows this is not a one-off. Another developer said GitHub acknowledged his June report about a fraudulent clone of his software and then went silent, until TinyURL moved quickly to disarm a virus-laden download for its users. A third linked a months-old post about a malware fork of the open-source Triton project still floating around. The pattern matters more than the individual takedowns: GitHub is now a search surface where imitation storefronts outrank the real vendor, the real-time consequences are full malware distribution, and the only effective escalation is the court of public opinion. It is the same trust gap our supply-chain trust boundaries and skill-file supply-chain articles keep circling: when distribution platforms assume the thing you built is safe, verification falls on the consumer.

Why it matters: "always download from the vendor" is no longer a courtesy warning - registration pages render on top of search results, and a vendor's only reliable lever to kill a squat repo is a Hacker News front page.

SECURITY

The UK's Two-Tier Encryption: Apple Users Split by a Missed Window

A MacAnorak essay (456 points, 406 comments) lays out the durable end state of the UK's secret order over Apple's Advanced Data Protection: two identical UK iPhone users now live under two different security regimes. One enabled ADP before Apple pulled the feature for new UK users in February 2025, after reported Technical Capability Notices asked it to build a capability to decrypt E2EE-protected iCloud. The other can never turn it on. Apple cannot roll ADP back for existing users by design - the setting is only changeable from the user's trusted devices - so the two-tier outcome is permanent until the underlying dispute resolves.

The dispute itself is now running its legal course in public. Apple lodged a fresh complaint at the Investigatory Powers Tribunal in July that became public in August. On September 11, Senator Ron Wyden and Congressman Warren Davidson wrote to the tribunal calling the secrecy order dangerous for both democracies, and on September 17 The Telegraph reported Apple asking to be allowed to confirm the notice's existence at all. The encryption-power argument touches whatever you build that touches user data, which is why we keep tracking it alongside agent security: our agent security comparison and yesterday's breach coverage are the same lesson from the other side of the fence - whoever can force access to your data system, user trust is the casualty.

Why it matters: for developers shipping E2EE or cloud-backup features, this is the template of the government tension you will face - and proof that the default resolution quietly reclassifies a privacy feature into a retroactive, irreversible privilege.

AI

Opus 5.5 Wrote, Directed, and Rendered an Explainer Video in One Prompt

LaunchVideo (270 points, 134 comments), built on the OpenComputer agent platform, turns "paste a URL or describe the product" into a finished MP4 with no edits: Opus 5.5 writes the film as an HTML page - layout, animation, copy - and a serverless agent renders it. The mechanics are the interesting part. The whole product is one TypeScript agent file with three tools (web_fetch, check_scene, render_video) running in a fresh microVM. No video model is involved: the page's clocks are replaced with a virtual clock so every frame is a deterministic seek, and the rendered HTML/CSS/JS is piped through headless Chromium into libx264 at 1920x1080, 30fps. The examples on the page - NVIDIA, Jev, Linear, OpenComputer - are each a single untouched run, about four minutes and roughly 100k tokens per film. The idea traces to Deedy's post about Opus 5.5 and instructional video: the model writes the film as code, and code renders the same every time.

This is the pattern we keep flagging from the agent-video lane: not model-generated pixels, but model-written, deterministic code that any renderer can play. It is closer to our auto-narrated changelog videos and OpenMontage agentic pipeline than to diffusion video, and it inherits the same economics we measured when GPT-6 Astra built a site: one long-horizon session produces a finished artifact that previously took a designer, a copywriter, an animator, and a video editor.

Why it matters: when a 100k-token run reliably produces a scoped, deterministic video, the unit cost of launch content collapses toward serverless compute - and the "writes it as code" trick is portable to any software artifact, not just film.

RESEARCH

Frontier Models Are Solving - and Re-asking - 17th-Century Questions

Historian Benjamin Breen's Res Obscura post (135 points) documents what GPT-6 and Opus 5.5 did when pointed at real historical problems rather than retrieval tasks. Opus 5.5 downloaded more than 5,000 files from the Samuel Hartlib archive, spun up sub-agents to cross-check unidentified sources across languages, and surfaced what Breen believes may be a new finding: both Newton and Hartlib used different anagrams for the key alchemical ingredient Hungarian vitriol (Newton's "vltimorui" is a true anagram of vitriolum; Hartlib's "Miloirtiua riciragnun" is imperfect reversed writing), which Breen reads as compelling evidence Hartlib's manuscript was the one Newton drew upon. Separately, Opus 5.5 partly deciphered two 16th-century Spanish letters in Emperor Charles V's secret code - and in a lesson about verification, both had already been cracked (one in the 1530s, one in 1916), which the model confirmed against the gold-standard plaintexts. GPT-6 Astra, meanwhile, concluded John Dee's "angelicall language" book is mostly nonsense syllables rather than code, and began an independence-style hunt through Darwin's notebooks for undetected links to his informants.

What is new is not transcription but hypothesis formation. The models proposed the research projects themselves, cross-checked multilingual archives, and reasoned about what would constitute a publishable result - the "spiky" capability profile they already apply to mathematics and cryptography. It is the intellectual neighbor of yesterday's Enigma break coverage and the Jacobian conjecture counterexample we recorded earlier, and Breen's conclusion is the practical one: the bottleneck is now access to digitized archives, not model capability, and he is arguing AI labs and funding agencies should start systematically pairing historians with frontier compute.

Why it matters: if frontier models can form and test historical hypotheses, the same toolchain powering code agents becomes a research instrument - and the verification discipline (checking against known plaintexts, acknowledging dead ends) is exactly what separates real findings from confident noise.

TOOLS WORTH A LOOK

1. Whiteboard (YC W26) (OSS, MIT) - an open-source desktop app where humans and agents architect software together, with an SDK for agents to draw on an in-app canvas, an AST-aware semantic diff viewer in Rust, and a decision log for tracing autonomous choices (313 points on HN).
2. Fearless SIMD v1.0 (OSS) - the Rust SIMD library from the Linebender team that statically generates operation sets, hits stable v1.0 with faster gathers, shuffles, and reductions across host and portable targets (262 points on HN).

WHAT ELSE IS HAPPENING

• Goodbye Google (192 points): Robert O'Callahan, creator of the rr reversible debugger and Pernosco, resigns from Google saying AI progress is "far too rapid," with his chip-design tooling work accelerating the exact capability he is worried about (he maintains rr and Pernosco either way).
• Dutch government builds an open alternative (161 points): a coalition of Dutch public bodies is standing up a NixOS-based replacement stack for Microsoft software, published as a community effort.
• Sourcehut account takeover via build logs (117 points): an XSS in the ansi2html renderer used on CI build logs lets an attacker steal a Sourcehut session and take over the account.
• Dynamic Abliteration (107 points): a non-destructive technique for suppressing a model's refusals via engram steering rather than weight surgery, pitched as a safer way to reason about safety filters.
• Datasette 1.0a41: Simon Willison's release adds OpenTelemetry support and refactors every modal dialog into one documented web component for plugins.
• New Jersey fines a data center $1.1M (85 points): drone and satellite imagery exposed 62 diesel generators the facility never disclosed, the kind of permitting reality AI-buildout runs into.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.