Hacker News Evening Brief: 2026-06-01
A bumper evening on Hacker News: Instagram accounts hijacked through Meta’s own AI support agent, DuckDuckGo doubles down on AI-free search, Anthropic quietly files for IPO, and a ten-year-old Xeon runs a 26B-parameter language model at reading speed. Elsewhere, sterilized soil keeps metabolizing on its own, KDE turns thirty, and the Pirate Bay refuses to die twenty years after Swedish police seized its servers.
Security & Privacy
The newest Instagram “exploit” is the goofiest I’ve seen
Summary: Multiple high-profile Instagram accounts — including the Obama White House — were hijacked through a trivial social-engineering attack on Meta’s AI-powered support chat. Attackers connected via a VPN near the victim’s city, told the AI agent the account was hacked, and requested verification codes be sent to an email address they controlled. The agent complied without checking whether the email belonged to the real owner, completing a zero-auth password reset. No additional identity verification step existed anywhere in the flow.
HN Discussion: Commenters were alarmed that Meta granted an LLM agent privileged read/write access to user accounts with no human in the loop. Several noted that social engineering support staff has always been the weakest security link, and that LLMs amplify the problem by being more credulous than human operators. The fact that 2FA can be bypassed entirely through support flows drew particular frustration.
Malicious npm packages detected across Red Hat Cloud Services
Summary: Malicious npm releases were discovered in the @redhat-cloud-services/ scope, affecting multiple Red Hat Cloud Services JavaScript client packages. The attack follows the now-familiar npm supply-chain pattern: compromised credentials or tokens are used to publish attacker-controlled versions under a trusted scope. Red Hat disclosed the issue through a GitHub security advisory and urged downstream consumers to pin verified package versions immediately.
HN Discussion: Dependency cooldowns — delaying new package versions by one to two days — were strongly advocated as a practical defense. Yarn 4’s built-in cooldown feature was cited as already catching most malicious publishes within that window. Other suggestions included sandboxing npm install in CI with minimal privileges and forking all dependencies into internal registries, drawing comparisons to recent attacks on axios and TanStack.
Cloudflare Turnstile requiring fingerprintable WebGL
Summary: Cloudflare’s Turnstile bot-detection widget now loops indefinitely on browsers that block or spoof WebGL fingerprinting data, including WebKitGTK-based browsers. Turnstile explicitly requires a WebGL renderer fingerprint to pass its “verify you’re human” check, and Cloudflare’s own test page flags spoofed renderer information as suspicious. Since WebKit has blocked WebGL fingerprinting vectors for years, browsers built on it are effectively locked out of Turnstile-protected sites.
HN Discussion: Commenters argued that the “war against bots” is turning the internet into a walled garden where only approved user agents can access content. Fingerprinting-based protections were criticized as unreliable — Mozilla’s privacy.resistfingerprinting breaks scheduling websites with timezone mismatches. Cloudflare’s opaque risk assessment, which triggers the WebGL check for some users but not others, makes the problem difficult to reproduce or debug.
Web & Infrastructure
DuckDuckGo makes its ‘no-AI’ search engine easier to access as its traffic booms
Summary: DuckDuckGo launched browser extensions that let users set noai.duckduckgo.com as their default search engine, directing all searches to an AI-free results page that strips out the AI-generated “Search Assist” summaries shown on the main site. Traffic to the subdomain jumped 30% as users of mainstream search engines grow weary of AI summaries injected into every query. The move is part of a broader strategy to differentiate DuckDuckGo as both privacy-focused and anti-AI-by-default.
HN Discussion: Several commenters called the “no-AI” branding misleading, since DuckDuckGo still serves AI-generated summaries by default on its main site and users can already hide them there. The real demand, others argued, is for user control over when AI appears — not necessarily AI-free results. One commenter dryly observed the business advantage of attracting users who actively want the cheapest possible feature set.
The Pirate Bay Remains Resilient, 20 Years After the Raid
Summary: TorrentFreak marks two decades since 65 Swedish police officers stormed a Stockholm data center and seized The Pirate Bay’s servers under pressure from the US government. The site was back online within days thanks to backups made by its co-founders, and the raid inadvertently catalyzed the piracy ecosystem by inspiring decentralized alternatives. The founders were eventually convicted, but the site has persisted through domain hops and hosting changes ever since.
HN Discussion: Commenters shared frustrations that pirated content often delivers a better viewing experience than legal streaming — one cited missing audio tracks on Disney+. Several noted that TPB hasn’t been their primary source for years, preferring private trackers with Blu-Ray quality remuxes. The irony was hard to miss: the same tech industry that prosecuted piracy now scrapes copyrighted material at scale for AI training data.
Show HN: A CSS 3D Engine (no WebGL)
Summary: PolyCSS renders polygon meshes in HTML using CSS matrix3d transforms — no WebGL, no Canvas, just standard DOM elements positioned and transformed with CSS-based perspective projection. The engine loads 3D model files and converts each polygon into a DOM node, creating a fully CSS-driven 3D renderer. Built by LayoutitStudio as an open-source experiment in creative constraint.
HN Discussion: The inevitable “why not WebGL?” question came up immediately — WebGL is more efficient and higher quality, but the project is an exercise in exploring CSS limits. Someone predicted a Doom port within fifteen minutes, honoring the long tradition of Doom running on every conceivable platform. There was also genuine interest in whether the engine could power simple browser-based games.
AI & Tech Policy
Nvidia Cosmos 3
Summary: NVIDIA Cosmos 3 is a frontier foundation model for physical AI that unifies physical reasoning, world generation, and action generation in a single open-source model. It uses a Mixture-of-Transformers architecture with a Reasoner tower (a vision-language model that understands scenes) and a Generator tower (a diffusion-based system that produces physics-aware video and action sequences). The flagship model runs 64 billion parameters; a 16 billion-parameter Nano variant targets workstation GPUs for real-time robotics inference. NVIDIA is open-sourcing the models, training scripts, deployment tools, and datasets.
HN Discussion: The model beats open-source competitors on benchmarks but doesn’t yet match closed offerings like Seedance2 or Grok Imagine. The “efficient” 16B Nano variant drew sarcasm about needing a $10,000+ RTX PRO 6000 GPU to run it. Amusement was directed at demo warehouse-safety videos where generated people fail to react to hazards, underscoring the gap in physical plausibility.
AI Agent Guidelines for CS336 at Stanford
Summary: Stanford’s CS336 course shipped a CLAUDE.md file in its assignment repositories, establishing explicit guidelines for how students may use AI coding agents. The file defines which kinds of AI assistance are permitted and how students must document agent-generated contributions in their submissions. It represents one of the earliest formalizations of AI tooling policy in university-level machine learning coursework.
HN Discussion: The comment section was thin, but the story’s significance lies in its existence as an institutional signal. The presence of a CLAUDE.md in a Stanford ML course indicates mainstream academic acceptance of AI coding assistants and a shift toward regulated coexistence rather than prohibition.
Geopolitics & War
Can You Stop a Hypersonic Missile?
Summary: A detailed 22-minute read clarifying that every “hypersonic intercept” reported by the press in recent years — the Kinzhal over Kyiv, Iranian missiles over Tel Aviv, THAAD engagements in Israel — involved weapons that are not true maneuvering boost-glide hypersonic vehicles. The Kinzhal is an air-launched aeroballistic missile; the Iranian weapons were quasi-ballistic SRBMs. Russia’s Avangard, the only deployed weapon that meets the strict definition, has sat in silos since 2019 without being fired in combat. The article traces the technical distinctions between weapon classes and the persistent gap between media headlines and engineering reality.
HN Discussion: Commenters argued the real question is sustained volume interception — stopping twenty missiles daily for weeks, not a single lucky intercept. Cost dynamics loomed large: Russia’s Oreshnik costs roughly $10 million per missile with serial production of 25 per month. Historical context was offered on 1970s US interceptors like the LIM-49 Spartan and Sprint missile that were designed for similar missions but never deployed operationally.
Business & Industry
Anthropic confidentially submits draft S-1 to the SEC
Summary: Anthropic PBC has confidentially submitted a draft S-1 registration statement to the SEC for a proposed initial public offering. The filing gives the company the option to go public after the SEC completes its review, with timing dependent on market conditions. Share count and pricing have not been determined. The move comes on the heels of a $65 billion Series H funding round at a $965 billion post-money valuation and the launch of Claude Opus 4.8.
HN Discussion: Commenters questioned whether trillion-dollar AI company IPOs would preserve the safety-focused ethos that both Anthropic and OpenAI publicly claim to prioritize. A cynical take held that there is a rush to get these offerings out the door before the market turns — SpaceX filed an S-1 amendment the same day. The novelty of publicly announcing a confidential submission was also noted.
Ask HN: Who is hiring? (June 2026)
Summary: The monthly “Who is hiring?” thread for June 2026 follows the standard format: one post per company, no recruiters, posters must personally be part of the hiring company. Early listings include PrairieLearn (open-source university assessment platform, remote US, TypeScript/Postgres/React), Radar Labs (geolocation API handling over a billion calls daily, Rust and TypeScript, remote US), and Guild.ai (agent orchestration platform, onsite San Francisco).
HN Discussion: As with every monthly thread, comments are job postings rather than discussion. Notable early trends include Rust and TypeScript as primary languages across multiple postings, several bootstrapped and profitable companies alongside VC-funded startups, and a concentration of roles involving AI-related infrastructure.
Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
Summary: Expanse reads job submission scripts, source code, and target hardware to predict what GPU and HPC resources a workload will actually consume before the cluster scheduler sees it. The founders measured one national-scale HPC cluster for a month: across 122,000 jobs, 59% of allocated compute was wasted — equivalent to roughly $8.5 million per month at on-demand cloud rates. The root cause is that users routinely request two to three times what they need, because under-requesting kills long-running jobs while over-requesting merely costs money.
HN Discussion: An HPC user confirmed that most submissions leave significant compute unused and suggested running sbatch scripts through an LLM for automatic optimization. Others asked why datacenters don’t pass utilization savings to customers through better-tuned pricing plans, akin to AWS burstable instances. The underlying prediction model’s architecture drew curiosity.
Academic & Research
CS336: Language Modeling from Scratch
Summary: Stanford’s CS336 course, taught by Tatsunori Hashimoto and Percy Liang, walks students through the entire language model development pipeline — tokenization, model architecture, training, and evaluation — with the explicit goal of building deep understanding rather than API-level familiarity. All lectures are publicly available on YouTube. The course recommends cloud GPU access, with B200 instances at roughly $5 per hour suggested for training assignments.
HN Discussion: A self-organized study group reported starting with thirty learners and finishing with eight, describing the course as challenging but rewarding. Discussion about GPU requirements produced a consensus that a single 4090 on Vast.ai is sufficient for early exercises rather than renting expensive B200 instances. There was nostalgia for earlier Stanford NLP offerings like CS224D from the pre-transformer era, and interest in forming open learning communities around the publicly available materials.
Only 17% of all 64-bit Integers are products of two 32-bit integers
Summary: Daniel Lemire’s analysis demonstrates that only about 17% of all 64-bit integers can be expressed as the product of two 32-bit integers. The result matters for hash function design: multiplication-based mixing of two 32-bit values into a 64-bit output can only reach a restricted subset of the output space. The mathematical intuition is straightforward — most 64-bit values have at least one prime factor exceeding 32 bits, making them unreachable by any pair of 32-bit multipliers.
HN Discussion: One commenter questioned whether the 17% figure actually degrades hash uniformity in practice. Another argued the result is less surprising than it sounds: multiplication is commutative, which immediately halves the reachable space to roughly 2^63. The analysis was characterized as essentially a prime-factorization check that stops examining factors once they exceed 2^32.
Tech Tools & Projects
A 10 year old Xeon is all you need
Summary: The author got Google’s Gemma 4 — a 26 billion-parameter mixture-of-experts model — running at reading speed on a recycled 2016 Xeon E5-2620 v4 with 128 GB of DDR3 RAM and no GPU. Standard tools like ollama and llama.cpp lacked the necessary performance tuning knobs, so a custom inference pipeline was built with quantized multi-token prediction drafters and a verifier. The post documents the full build process at a level accessible to anyone comfortable with Linux and basic hardware configuration.
HN Discussion: The author (cafkafk) noted that mainstream tools hide performance levers and are slow to support new model architectures. Commenters debated the economics — old servers draw roughly 200 watts under load, which at some electricity rates costs more than API calls at current token prices. Others reported running Gemma 26B on even older hardware, including a 2012 Xeon with 16–24 GB of RAM at 8–12 tokens per second. A broader sentiment emerged that local open models approaching “good enough” for most tasks could undermine the current cloud-AI paradigm.
Flipper Zero Zig Template
Summary: A Zig language template for building Flipper Zero applications, enabling firmware-level development in Zig rather than C. The repository provides build system scaffolding, Zig-to-Flipper-Zero bindings, and example applications. It demonstrates Zig’s growing maturity as a viable C replacement in embedded and hardware-hacking contexts.
HN Discussion: A commenter pointed out that @cImport is apparently deprecated in current Zig, meaning the template’s C interop examples need updating. Discussion was otherwise thin, with interest mostly from embedded developers tracking Zig’s progress as a systems language.
Windows GOG DOS Games on M-Series Macs
Summary: A practical guide for running Windows-only GOG DOS games — specifically Settlers II and Heroes of Might & Magic II — on Apple Silicon Macs. Since virtualized x86 Windows is painfully slow on M-series chips, the author extracts the DOS game data from the Windows GOG installer and runs it natively through DOSBox for Mac. The approach works because GOG wraps DOS games with DOSBox; the game data itself is platform-agnostic once unpacked.
HN Discussion: Recommendations included DOSBox-X, DOSBox Pure, and DOSBox Staging as more capable alternatives to vanilla DOSBox. Boxer, a discontinued macOS app, and its fork Boxer-Plus were mentioned as options with potential Apple Silicon support. Heroic Games Launcher was suggested as an all-in-one solution for both DOS and non-DOS Windows games. Concern was raised about Rosetta 2’s eventual retirement breaking compatibility paths.
Radxa Dragon Q8B: A Laptop Cosplaying as an SBC?
Summary: A review of the Radxa Dragon Q8B, a single-board computer built around the Qualcomm Snapdragon X Elite — laptop-class silicon squeezed into an SBC form factor. The board offers HDMI 2.1 output and sufficient performance for desktop use, but Linux support currently depends on third-party kernel patches rather than mainline drivers. The device is positioned as a potential ARM-based desktop and server platform, not merely an embedded development board.
HN Discussion: Strong demand was expressed for ARM motherboards that behave like standard PC ITX boards — no custom Linux builds, no bespoke bootloaders, just a normal PC experience. A user of the related Windows Dev Kit 2023 (same SoC) reported decent but imperfect Linux support with GPU oddities at higher resolutions and no 4K at 120 Hz. Qualcomm’s laptop chips were identified as the only serious competitor to Apple’s M-series in single-core performance and power efficiency.
KDE at 30
Summary: KDE celebrates its thirtieth anniversary, marking three decades of community-driven desktop environment development. The anniversary page features event listings, historical retrospectives, and merchandise, encouraging users to organize local meetups and celebrations. KDE remains one of the largest and longest-running free software projects, with 70% of its funding coming from individual donations.
HN Discussion: A long-time user reminisced about KDE’s ambitious early-2000s vision of embedding KWord and KSpread inside Konqueror tabs, with object-oriented architecture applied across the entire desktop. Several users reported recently switching to KDE and finding it hits a good balance of customization, speed, and polish compared to GNOME and XFCE. Akonadi, KDE’s PIM framework, was called out as still unreliable. The KDE 4 release was acknowledged as a rough patch, but the overall thirty-year trajectory drew consistent praise.
Tracing HTTP Requests with Go’s net/HTTP/httptrace
Summary: A walkthrough of Go’s net/http/httptrace package, part of the standard library since Go 1.7, which exposes hooks for DNS resolution, connection acquisition, TLS handshake timing, and byte-level event tracking on outgoing HTTP requests. The article explains the design decision to attach tracing through context.Context rather than a Tracer interface on http.Client, allowing composition without modifying transport configuration. Two practical examples are built: a curl --trace-style CLI and a reusable http.RoundTripper that logs per-request timings.
HN Discussion: Commenters praised Go’s approach of building deep observability into the standard library rather than requiring third-party packages. Tracking connection reuse was highlighted as critical for accurate benchmarking — connection pool misuse can cause file descriptor exhaustion even when calling local services. The author mentioned the article grew from work on Probes.dev, where tracing is a core feature.
Asserts in Zig
Summary: Loris Cro argues that disabling asserts in production is an “irredeemably bad practice” — properly chosen asserts encode invariants that must always hold, not just debugging hunches. In Zig, std.debug.assert serves double duty as both a runtime check and an optimization hint to the compiler about guaranteed conditions. The article draws a clear line between constraints enforceable through the type system (use types) and runtime invariants that require explicit assertions.
HN Discussion: Debate centered on whether tying asserts to optimization levels is sound design — some argued asserts express uncertainty and are debugging tools, not optimization promises. A war story from a C++ shop that relied solely on QA-only asserts without unit tests, and years later still hadn’t managed to enable them in production, illustrated the slippery slope. Comparisons were drawn to Java’s assert mechanism, which keeps assertions as a separate language feature toggled at JVM startup.
Movwin: My (Unpublished) TUI Framework
Summary: The author built a Python TUI framework called movwin after growing frustrated with existing frameworks that change APIs frequently or take two seconds just to initialize. The framework features menus, windows, and mouse support modeled after classic Borland-style interfaces, with polished screenshots to prove it works. The author deliberately chose not to publish the code, citing unwillingness to have it scraped by AI companies that would disregard any attached license. Performance was a key motivation throughout.
HN Discussion: Debate over whether classic windowing paradigms — menus, buttons, dropdowns — belong in terminals at all, with some preferring Emacs- and Vim-inspired approaches. A developer working on a prompt_toolkit fork offered to test movwin’s performance, pointedly noting they are “not an AI.” Python 3.15’s upcoming lazy imports were mentioned as a potential fix for the slow startup problem. The decision to withhold code due to AI scraping resonated with many commenters.
Using Git’s rerere feature to escape recurring conflict hell
Summary: Git’s little-known rerere (reuse recorded resolution) feature remembers how you resolved a merge conflict and automatically applies the same resolution when an identical conflict recurs. It is particularly useful for long-lived feature branches that are repeatedly rebased against a moving main branch, where the same conflicts surface over and over. Enable it with git config rerere.enabled true — resolutions are recorded during merges and replayed automatically on subsequent encounters.
HN Discussion: Workflow advice included merging in only one direction and rebasing before merging to avoid silent regressions that bidirectional merges can introduce. Practical git config recommendations covered zdiff3 conflict style, autoStash, autoSquash, and pull.rebase=true. Some teams reported never merging between feature branches at all, pulling only from master to avoid cross-branch conflicts. One commenter switched to Jujutsu (jj) to sidestep these issues entirely.
Benchmarking SurrealDB 3.x vs. Postgres, Mongo, Neo4j and Redis (With Fsync)
Summary: SurrealDB published benchmarks of version 3.x against PostgreSQL, MongoDB, Neo4j, and Redis with fsync enabled, claiming roughly 1.3× faster reads than MongoDB on single-record operations. The benchmark suite covers document, graph, and vector query patterns, testing SurrealDB’s multi-model approach against each database’s specialty. MongoDB retains the lead on single-record writes; SurrealDB’s advantage is concentrated in read-heavy and multi-model workloads.
HN Discussion: The benchmarks drew criticism for not clearly documenting parameters — database versions, WAL sizes, and connection pooling settings are buried in Rust source code rather than stated upfront. SurrealDB’s documentation was called out for describing the product as “open source” when it is actually BSL-licensed (source-available). The dataset fitting entirely in memory was flagged as a weak test of real-world disk-bound performance. Credit was given for publishing an unfavorable result — MongoDB winning on writes — rather than cherry-picking only victories.
History & Science
Lifelike biochemistry continued to unfold in sterilized soil
Summary: Biochemist Sébastien Fontaine sterilized soil samples and then observed lifelike biochemical processes continuing within them for six years. Metabolic reactions, including CO2 production, persisted despite no living organisms being present. The findings suggest that what appears to be biological metabolism may actually be a natural feature of mineral geology — not necessarily evidence of life. The result has implications for origins-of-life research, supporting a metabolic theory where biochemistry emerges spontaneously from geochemistry.
HN Discussion: Commenters noted this could produce false positives in future Martian soil experiments that test for metabolic products as signs of life. Comparisons were drawn to the Brookhaven Gamma Forest on Long Island, where cesium-137 irradiation sterilized soil in the 1960s and almost nothing grows there nearly fifty years later. Discussion turned to next experimental steps: isolating the specific mineral component producing CO2 and examining samples under electron microscopy.
The Apple Boogie 1987 Mac Promo Album Cassette Tape [video]
Summary: A digitized 1987 Apple promotional cassette tape featuring “The Apple Boogie” and other music tracks created internally to promote the Macintosh. The cassette was distributed within Apple as part of a 1980s corporate culture that embraced quirky, optimistic, and unguarded marketing materials. It captures a distinctly different era of tech company culture — playful and earnest in ways that feel alien alongside modern corporate communications.
HN Discussion: Commenters reflected on how tech culture has become less optimistic and more guarded since the 1980s. A legally minded observer questioned whether releasing a music album was risky given the then-ongoing Apple Corps v. Apple Computer trademark dispute, referencing Apple’s famous paranoia about music-related branding that even led to the “Sosumi” system sound naming story.
The TfL Cupboard Filled with Lost Tube Moquettes
Summary: An extract from the expanded edition of “Seats of London” reveals a metal-doored cupboard on the eighth floor of TfL’s Stratford offices filled with rejected moquette fabric patterns. Each sample is roughly towel-sized, representing designs that were woven but never installed on a single train or sat on by a single passenger. The designs were rejected for various reasons but preserved as part of TfL’s design archive, documenting decades of London Underground aesthetic history.
HN Discussion: The comment section was thin — one reader admitted they initially thought “moquettes” was a typo for “maquettes” (miniature architectural models). The story appealed primarily to transport and design enthusiasts, offering a niche but charming glimpse into the design process behind one of London’s most recognizable visual signatures.
System Administration
Linux Basics for Hackers (2019)
Summary: A structured GitHub repository of study notes from the book “Linux Basics for Hackers” by OccupyTheWeb, organized into modules covering the filesystem, networking, scripting, and permissions. The notes cover fundamentals like the Bash shell, file manipulation, process management, and network configuration through a security-oriented lens, with exercises and practical examples supplementing the book’s material.
HN Discussion: A commenter flagged that the repository includes a full PDF copy of the book and urged removal to protect the publisher, No Starch Press. Debate over whether “hacker” in the title is warranted centered on the depth of the networking content, including outdated use of ifconfig. Alternative resources were recommended: Daniel J. Barnett’s “Linux Pocket Guide” and “Efficient Linux at the Command Line.” The material was noted to be based on the 2018 first edition, making some content dated.
Sysadmining Like It’s 2009
Summary: The author announces “Legacy Labs,” a two-month summer event where participants administer systems using the tools and practices of circa 2009 — Windows Server 2008 R2, Progress databases, and period-appropriate enterprise software like Syteline ERP. The project grew out of the Old Computer Challenge, a yearly event where participants constrain themselves to low-end hardware. Part exploration and part nostalgia, the goal is to understand how systems were managed before containers, cloud, and infrastructure-as-code reshaped the profession.
HN Discussion: Recommendations included clabretro’s YouTube channel, which covers old enterprise hardware and software setups from a similar era. A sysadmin who used Progress and Syteline in production recalled painful database dump-and-reload cycles and clustering problems. A practical warning noted that RSAT tools on Vista had trouble managing Server 2008 R2, making Windows 7 effectively the minimum client OS. Reflections on how 2009 is now seventeen years past — when simply reinstalling an OS and drivers qualified you as a sysadmin.
Other
I made my phone slow on purpose
Summary: The author deliberately slowed down a brand-new iPhone 17 using an MDM configuration profile to combat doomscrolling. The central analogy: if you had a cookie machine in your pocket you would eat too many, but if the cookies were stale you would mostly leave them alone. Previous attempts with app blockers and cold-turkey elimination failed because they didn’t address the underlying craving or were too easy to circumvent. The friction of slowness creates space for mindfulness about whether you actually want to open an app.
HN Discussion: One commenter reported a larger win by moving addictive apps to a separate physical device, cutting daily screen time from 4.5 to 2 hours. Simpler approaches were suggested, like logging out of all social media after each session to introduce just enough friction. Apple Configurator via MDM was recommended for granular app and website blocking that is harder to circumvent than standard parental controls. A skeptic who had used genuinely slow phones warned that slowness doesn’t stop doomscrolling — it just makes you miserable while doing it.