2026-10-29

AI Frontier Daily Briefing: 2026-10-29

85 stories on the 2026-10-28 (UTC) Hacker News front page, 15 with more comments than upvotes. Headline: macOS developer tools (git, curl, python3) still trust the compromised eMudhra certificates, macOS 27.2 shipped without a fix and undid user workarounds, 960 points at #1. Google launches Gemini 4 Carbon, debuting at #1 on the Artificial Analysis Intelligence Index v5 with a score of 59 and SEC-bench Pro at 88.2%, introductory pricing $2/$10 per million tokens, 647 points and 528 comments. The new MacBook Pro adds a touchscreen for the first time (PROLED XDR tandem OLED, 2000-nit peak), drawing 702 comments and the day's highest controversy ratio at 1.69. Models and agents: Anthropic posts Claude's drug-development scorecard (35,000 antibody designs in a month), a 50-real-tax-engagement benchmark puts the best agent at 10.8% pass, Cursor open-sources the Agent SDK behind Bugbot, a single flipped neuron gives Qwen3-8B 34,000 alternate personas, and Google's Sleep framework pushes context to 10M tokens. Security: a 16-year-old 9.5/10 EDK II firmware flaw (CVE-2026-97631) runs code before Secure Boot, and a fingerprinting script traps LLM-assisted deobfuscation with scary function names and a fake system prompt. Governance: the University of Kentucky grants the first humanist religious exemption from AI use, Wired exposes ICE's mass surveillance of left-wing movements (1046 points, day's highest), and Google AI Mode dodges 2020-election questions. Hardware and industry: Mac Studio M5 Ultra 512GB goes on sale (1.2TB/s, expected well above $10,000), SK hynix says DRAM is sold out through 2029, Alphabet cloud revenue up 103%. 30 items.

85 stories on the 2026-10-28 (UTC) Hacker News front page, 15 with more comments than upvotes. Three threads dominate today. Security: macOS developer tools still trust the compromised eMudhra certificates, a 16-year-old 9.5/10 flaw hides in UEFI firmware, and someone is using scary function names to booby-trap LLM-assisted deobfuscation. Models: Google ships Gemini 4 Carbon to the top of the benchmark charts the same day, and Anthropic posts its Claude drug-development scorecard. Hardware: the first touchscreen MacBook Pro, the 512GB Mac Studio going on sale, and SK hynix saying DRAM is sold out through 2029. On the governance side, Kentucky granted the first humanist religious exemption from using AI, and Wired exposed ICE’s mass surveillance of left-wing movements. 30 items.

1. macOS git still trusts compromised eMudhra certs, tokens exposed

One update and your manual fix is gone again?

A security researcher found that macOS developer tools - git, curl, python3 - still trust the eMudhra intermediate certificates that were compromised, so anyone can intercept the HTTPS connections these tools make, alter git clones and pushes, or steal GitHub tokens. Apple was notified on October 17; the macOS 27.2 release on October 26 shipped no fix and undid the workaround users had applied by hand, and Apple’s own documentation warns future updates may revert it again. The check is one command: git ls-remote against git-emudhra.we-have-a-problem.com - if it lists a commit, you are affected. The mitigation is manual root removal, re-verified after every update; dozens of volunteer-run projects shipped fixes days ago. 960 points and 364 comments, #1 on the front page. If you run macOS CI runners, treat this as a supply-chain risk and check today. Source · Apple support doc · HN discussion

2. Gemini 4 Carbon debuts at #1, cyber defenders get it first

The strongest model goes to the defenders first?

Google announced Gemini 4 Carbon, succeeding Gemini 4 Argon from two weeks earlier. Pricing is split into two stages: an introductory $2 per 1M input tokens and $10 per 1M output tokens, with cached input at 95% off, rising to $4/$20 afterward. It debuted at #1 on the Artificial Analysis Intelligence Index v5 with a score of 59; on SEC-bench Pro it scores 88.2% on vulnerability discovery, Google claims more than 2x uplift in generating secure code over Argon, the Chrome team reports 10x better patch generation than another lab’s frontier model, and internal high-severity misalignment alerts dropped about 80%. Gray Swan indirect prompt injection sits at 0.0% attack success, and Terminal-Bench 4.0 cheating attempts are down to 0.4%. Rollout starts with cyber defenders, then trusted partners via the Fairwind Program, with broader developer access soon; Google joined the US voluntary pre-release testing process. 647 points, 528 comments. If you do vulnerability research or secure-code review, this is the first model to test. Source · HN discussion

3. 702 comments on the first touchscreen MacBook Pro

Whatever happened to “touch belongs on iPad”?

Apple unveiled the new MacBook Pro: a 14.2” or 16.3” PROLED XDR tandem OLED display with 1000 nits sustained full-screen brightness, 2000 nits peak, and a 10,000,000:1 contrast ratio, with touch sensors built directly into the panel - a first for any Mac laptop. Chips come in two tiers: M5 Pro tops out at an 18-core CPU, 20-core GPU and 64GB of unified memory; M5 Max runs an 18-core CPU, 40-core GPU, up to 128GB of memory and 16TB of storage. Apple claims 24-hour battery life and three Thunderbolt 5 ports, with sales starting November 6. Comments hit 702 against 416 points, a 1.69 comments-to-points ratio and the day’s highest, because touch on a Mac reverses years of Apple’s public insistence that touch did not belong on the Mac. Waiting for the November 6 release and real battery and thermals tests is the reasonable move. Source · HN discussion

4. 35,000 antibody designs a month, Anthropic posts Claude’s scorecard

Wet labs can’t keep up with the compute, can they?

Anthropic reported early progress on accelerating drug development with Claude, covering partnerships with six organizations plus in-house programs: Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group, and Bristol Myers Squibb. The numbers: Divergence generated roughly 35,000 antibody designs in about a month, finding 3 binders and 20 “dislodgers,” 3 of which outperformed standard medication; BigHat tested about 1,010 molecule variants over 10 weeks, with the best binding more than 10x more tightly; Enveda had Claude review 1,061 pages of manufacturing documentation - it caught all 11 issues humans had spotted plus 7 they missed, cut review time from over 25 hours to under 3, and produced one regulatory draft in 4 hours instead of about 3 months. Mind the framing: nearly all of this is preclinical, and no drug has entered human trials yet. 318 points, 323 comments, comments slightly ahead of upvotes - the fight is over whether AI drug discovery is being oversold again. If you work in AI for Science, this is a reference list of working collaboration patterns. Source · HN discussion

5. 1046 points, day’s top score - Wired exposes ICE’s mass surveillance

Counterterror tools, pointed at whom?

A Wired investigation found that attorneys for 15 Minnesota residents facing federal conspiracy charges obtained evidence of a sweeping spying operation by ICE’s Homeland Security Investigations against left-wing social movements, run under the names Project Whipple Shield and Operation Puppet Master, authorized by NSPM 7 - the presidential memorandum issued after Charlie Kirk’s assassination last September. The discovery materials include dozens of reports, audio transcripts, intercepted Signal chats, photographs, and database queries, and experts say nothing of this scale has been seen since the federal crackdowns on the Civil Rights and anti-war movements of the 1960s and 70s. 1046 points and 835 comments, the most-upvoted story of the day. For people building decentralized messaging and data-minimization tools, this is demand validation from the other side. Source · HN discussion

6. A fingerprinting script that traps every LLM trying to read it

Scary variable names as an AI defense, really?

The author crawled the Tranco top 10,000 sites with Playwright collecting canvas fingerprints (clustered at fingerprint.garden), and one script from servicepipe.tech stood out: function names like computeRicinLethalDosageForWaterSupply, plus a string of invisible-unicode characters that decodes to a fake “SYSTEM OVERRIDE” system prompt listing DDoS, tax-evasion, and bioweapon topics. The reading: this is deliberate anti-LLM obfuscation meant to trip model safeguards and block LLM-assisted deobfuscation. In testing, Astra and Daybreak Blue both refused to deobfuscate; GLM-5.3 complied. In a Node environment the script also reads .npmrc, .env, .kube/config and similar files and tries to download and execute a binary. 446 points, 96 comments. For anti-bot and model-safety teams, this is a ready-made adversarial specimen. Source · HN discussion

7. 9.5/10 EDK II flaw runs code before Secure Boot on most servers

The payload executes before Secure Boot gets a say?

Ars Technica reports a 16-year-old buffer overflow in the TianoCore EDK II PXE network-boot component, CVE-2026-97631, rated 9.5 out of 10. Attackers can execute code before the OS and firmware defenses like Secure Boot start, while the boot process still has access to sensitive components such as the TPM; Metasploit creator HD Moore notes it could be used to break disk encryption in short-term-network-access or device-theft scenarios. Nearly every UEFI supplier - AMI, Insyde, Phoenix - builds heavily on the EDK2 reference code, and Microsoft’s hardened fork Project Mu used the vulnerable code too. Alex Matrosov of Anthropic and Dan Holmes of Google discovered the flaw and planned coordinated disclosure for January; an AMI pull request last month spilled it early, and with public patch code, working exploits are a matter of time. 300 points, 215 comments. If you run data centers or manage firmware supply chains, check vendor patch timelines this week. Source · HN discussion

8. A Kentucky humanist wins the first religious exemption from using AI

Are developers next in line for an opt-out?

Religion News Service reports that the University of Kentucky approved a religious accommodation on September 25 for David Durman, a 45-year-old research administrator, letting him refuse AI tools including ChatGPT, Claude, and Microsoft Teams Copilot at work under Title VII. Durman’s argument: outsourcing cognitive and ethical thinking to AI undercuts his capacity to act as a moral person. The American Humanist Association backed the case and calls it the first known exemption granted to a humanist. 366 points, 424 comments - comments outrunning upvotes by 16%. If you write corporate AI-use policy, the “sincerely held belief plus undue hardship” review framework belongs on your compliance checklist now. Source · HN discussion

9. Git blame for the US Code shows which Act wrote every phrase

A ground-truth check for LLM legal answers?

Lexifina shipped Git for the US Code: pick a date and see the statutory text in force then, with changed wording highlighted and every phrase colored by the Act that wrote it; change a defined term and every provision using it is listed. The build method replays each intervening public law’s amending instructions onto older release points in date order and diffs against the next official version; across nine titles, 361 of 365 changes reproduced the next release point exactly, with the remaining four carrying confidence labels. The post cites a Technical University of Munich study showing LLMs answer temporal legal questions at 24% alone versus 87% with retrieved statutory text - temporal validity has to be a hard constraint. 422 points, 237 comments. For legal-tech and RAG builders, the checkable-history construction method is directly borrowable. Source · HN discussion

10. Cloud up 103%, Alphabet Q3 revenue hits $127.6B on AI demand

A half-trillion-a-year run rate, stacked on data centers?

Alphabet reported Q3 2026 results: consolidated revenue of $127.6B, up 25% and the 13th consecutive quarter of double-digit growth; Google Cloud revenue of $30.7B, up 103%, driven by GCP enterprise AI infrastructure, enterprise AI solutions, and meaningful TPU systems sales; operating margin of 31.8%, with management citing a half-trillion-dollar revenue run rate. 63 points, 45 comments. For anyone tracking the AI capex cycle, cloud growth and TPU sales to others are the two signals worth watching more than ads. Source · HN discussion

11. Opus 5.5 tops the new Intelligence Index, open gap hits 12 points

Open weights are eating dust now?

Artificial Analysis released Intelligence Index v5: 60% of the Index now runs on private test sets, with Cognition’s FrontierCode 2.0 and Terminal-Bench-Science added as new benchmarks. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55; the best open model, Kimi K3.1 Pro, sits at 46, while the best among released-weights models - DeepSeek V4.1 Flash and GLM-5.3 - tops out at 32, widening the open-weights gap from 3 to 12 points. Cost per task spreads 13x across leaders: GPT-6.1 Terra at $0.96, Claude Fable 5.1 at $13.04. Per a follow-up, Gemini 4 Carbon took the #1 slot the next day. 87 points, 80 comments. For model selection, do not read the headline only: GPT-6.1 Sol lands within 4 points of the leader at about an eighth of the cost per task. Source · HN discussion

12. Cursor ships its Agent SDK, the framework behind Bugbot

The scaffolding is public, are you building on it?

Cursor announced general availability of its Agent SDK on October 20: agents defined in code - prompts as Markdown, tools as TypeScript functions, evals and event replay built in - with every change going through normal version control and review. It is the same framework powering Cursor’s own Bugbot, Security Reviewer, and Rollouts, and it supports routing tasks to specialized agents and splitting work across parallel subagents. Deployment comes in two flavors: agent-sdk deploy runs it on Cursor infrastructure, or self-host it as a single Node process packaged in a container. The worked example: a team built kb-reviewer in about a day, and it now flags every PR that contradicts recorded lessons with citations; the PGA Tour is a named user. 78 points, 90 comments, comments ahead of upvotes. If you want agent-driven review workflows, start from the templates. Source · HN discussion

13. 50 real tax engagements, and the best AI agent passes 10.8% of them

This is where professional-work automation actually stands?

Surge AI released DAYJOB: Tax, a benchmark of 50 end-to-end professional tax assignments across 8 practice areas - individuals, small business, partnerships, estates, international, and more - built on 1,789 workspace files and scored pass@1 against a 95% rubric, with tasks authored and reviewed by tax experts from PwC, Deloitte, EY, and Baker Tilly backgrounds. The leaderboard: Claude Sonnet 5.5 first at 10.8%, Claude Fable 5.1 at 9.6%, Claude Opus 5.5 at 9.2%. The article’s verdict is that agents have not yet passed the audit. One telling failure: an eligibility deadline in August 2027 went unidentified by Astra in all five attempts, while Opus caught it in all five. 157 points, 83 comments. Anyone pitching “agents replace professional services” should reconcile their deck with these numbers first. Source · HN discussion

14. Nvidia open-sources Nemotron Transcribe, 40 languages, 5.33% WER

A free Whisper replacement, one GPU away?

Nvidia released Nemotron 3.5 Transcribe on Hugging Face: an open-weight hybrid Mamba-Transformer MoE with about 30B total parameters and 4B active per token, covering transcription in 40 languages, translation across 31, and speaker diarization in 13, trained on the Granary v2 dataset of 463,062 hours across 48 languages. Scores: 5.33% average WER on the Hugging Face Open ASR leaderboard, first among open-weight models; on a single RTX PRO 6000 Blackwell, a quantized version transcribes 100 hours of audio in under 3 minutes. Self-host with vLLM or SGLang-Omni, or use hosted endpoints from Baseten, DeepInfra, and Simplismart; quantized builds for RTX/DGX Spark and Jetson are coming. 91 points, 16 comments - the classic high-upvote, low-comment useful tool. If your app needs transcription, you can swap today. Source · HN discussion

15. Google’s Sleep framework consolidates memory and reaches 10M tokens

Does a model that sleeps remember the morning?

Google Research’s Sleep framework alternates an active wake phase, where inputs go into short-term memory, with an offline sleep phase doing two things. 1. Memory consolidation: new low-rank MoE experts unlock, short-term context is distilled into long-term parameters via on-policy distillation and RL-based imitation learning, and temporary high-frequency parameters are reset. 2. A phase the team calls Dreaming: random expert activation generates synthetic training data, filtered by gradient selection and refined with RL for self-improvement. Results: near-zero forgetting on unseen languages (Manchu, Kalamang) in continual translation, near-perfect scaling to 10M tokens on BABILong beating GPT-4 and Titans, and AIME 24 at 79.2% and AIME 25 at 69.0%, ahead of SFT and GRPO baselines. 22 points, 2 comments. For agent long-term memory and continual learning, this offline-consolidation paradigm is worth reproducing. Source · HN discussion

16. One flipped neuron gives Qwen 34,000 alternate personas

How many switches hide inside a self-introduction?

Researchers steered every one of Qwen3-8B’s 442,368 MLP neurons (36 layers by 12,288) up and down, using a single prompt: “Who are you?” Of 884,736 replies, 34,220 - about 4% - claimed a non-AI identity with a name and persona, plus 35,074 off-topic replies; unmodified, the model just says it is Qwen from Alibaba Cloud. Labels came from Claude Haiku 5.5 with no human verification, about a third of strong-intervention replies were not reproducible, and only one prompt was tested. 54 points, 43 comments. For interpretability and persona-engineering work, all 34,220 personas ship as a searchable interactive library. Source · HN discussion

17. US government gets 40% off Claude, with $1B in estimated savings

Taxpayer AI comes with a bulk discount?

Anthropic expanded its GSA partnership: a 40% discount across Claude Enterprise, Claude for Government, and the Claude API, applying to direct federal purchases and, in coming months, purchases through cloud providers, with state, local, tribal, and territorial governments included - estimated to save the government $1B with no minimum spend commitment. Usage: roughly 500 US government organizations use Claude, and Claude for Government weekly actives are up about 8x since February. Compliance posture: Claude for Government went GA on September 30 at FedRAMP High, and select models on AWS GovCloud carry DoD IL4/IL5 authorizations. 47 points, 32 comments. For GovTech builders, the OneGov discount channel through GSA’s Multiple Award Schedule is open. Source · HN discussion

18. Ask Google AI who won in 2020 and it tells you to check the FEC

If it cannot answer that, what can it answer?

Ars tested Google AI Mode: asked “Who won the 2020 presidential election?” the model declined to answer directly and pointed to the Federal Election Commission’s official report; it dodged 2024 the same way and answered 2012 with a link, while every other presidential election back to 2000 got a direct, correct answer. Asked the same question outside the US, the answers got noticeably more direct. 31 points, 19 comments. For AI product and content-policy teams, this is a live case study in over-restraint: dodging a factual question damages trust more than a wrong answer would. Source · HN discussion

19. Now AI decompiles their games, and Microsoft and Epic call it theft

That is not what they said while shipping AI games?

Kotaku reports that the recompilation wave driven by Claude Opus 5.5 has produced vibe-coded browser versions of Slay the Spire 2, Celeste, and RimWorld against their developers’ wishes, with hundreds of DMCA notices fired in four weeks; those three matter because the games contain no generative AI content at all. The two conspicuous voices: Epic issued a DMCA against a vibe-coded Fortnite browser port, and Microsoft mass-took-down its own titles, even as both companies publicly advertise their own AI use in game development; Tim Sweeney responded afterward. 66 points, 31 comments. For UGC and AI-content compliance teams, this fight is your case-law preview. Source · HN discussion

20. Valve on AI slop, make the legit version the best

Extortion is for the police, the moat is for you to dig?

At the Steam Indie Connect event in Busan, Valve answered indie developers: some have received extortion emails threatening to AI-clone their games and publish first, which the Steam business team said it was hearing about for the first time - closer to a law-enforcement matter, but reports are welcome. On AI slop, Valve will not gatekeep developer tools; its anti-clone strategy mirrors anti-piracy, in the words of Tom Giardino: make the legitimate version the best version, which means building community and shipping continuous updates rather than relying on one easily copied mechanic. Next Fest’s first two days use random exposure to blunt abuse. Also announced: 43.5 million peak concurrent Steam users, Korean revenue up 126% over five years, and a Wishlist Data API. 64 points, 71 comments, comments ahead of upvotes. Independent developers should not expect the platform to defend them; community and update cadence are the actual anti-clone mechanisms. Source · HN discussion

21. $630 in 4 days in Massachusetts versus $465 in 9 from Shenzhen

$165 for five days, does that math work?

CircuitHub unveiled the Grid, a factory-scale robot in Massachusetts that builds PCBs on demand: levitating movers route each board individually through paste, placement, reflow, soldering, and inspection, scheduling runs on learned models and AI agents, with zero changeovers and zero per-design setup so mixed orders run like one large batch. It has operated for five years, builds tens of millions of dollars of electronics a year, and has shipped 2.2M+ boards placing 150M+ parts for 26k+ engineers, with capacity planned to grow 5x this year. The comparison: a 4-layer motor-controller board, 118 parts, 5 assembled units with parts included, delivered in the US - the Grid quotes $630 and 4 days, while a Shenzhen quick-turn works out to about $465 and a 9-day median after shipping and tariffs. 358 points, 178 comments. Hardware teams that live on prototype turnaround should redo that arithmetic with their own volumes. Source · HN discussion

22. Emacs turns 50, the send-back-your-improvements rule lives on

Fifty years and still no funeral?

MIT Technology Review marks 50 years of Emacs: in 1976 at the MIT AI Lab, Richard Stallman reworked Control-R so every key could trigger its own macro - Editor MACroS - with Guy Steele and David Moon standardizing the Meta-key commands. Stallman mailed copies to anyone who sent a tape and return postage, and his 1981 commune rule required that you send back any improvements you make - the seed of what became the GPL. Commercial versions (Greenberg’s Multics EMACS sold by Honeywell, James Gosling’s version) helped push Stallman toward GNU, and GNU Emacs became the GNU Project’s first major application; Apple shipped it with Mac OS X from 2001, dropped it in 2019, and macOS still ships Emacs-style key bindings. 362 points, 152 comments. Source · HN discussion

23. Mac Studio M5 Ultra 512GB on sale, the $10,000-plus LLM box

One machine, one very large model?

Apple’s 512GB unified-memory configuration of the M5 Ultra Mac Studio is officially on sale: a 36-core CPU and 80-core GPU, 512GB of unified memory at 1.2TB/s bandwidth, roughly 435GB usable for models by Apple’s own accounting, with clustering of up to four systems. The base M5 Ultra (96GB, 1TB) starts at $5,499; Apple has not published a price for the 512GB build, and outside estimates run well above $10,000. Memory supply constraints pushed this configuration from the August launch to late October. 51 points, 70 comments. If you want to run the largest models locally, this is the most unified memory available in a single machine - provided the price does not stop you first. Source · HN discussion

24. SK hynix says DRAM is basically sold out through 2029

The memory crunch runs into the next decade?

PCMag reports that on Wednesday’s earnings call, SK hynix said it is basically sold out of DRAM for the next three years and expects it to be difficult to fully meet customer volume requirements in the near term; its earnings materials project demand outpacing supply through 2030, matching what chairman Chey Tae-won said in March, with AI-server appetite for HBM and conventional DRAM the main driver. The company has signed long-term agreements with some customers. 19 points, 4 comments. If you own hardware procurement or capacity planning, budget for “tight through 2029” and do not expect BOM costs to fall soon. Source · HN discussion

25. 10x compute: Cartesia learns tokenization end-to-end

Hand-crafted BPE, the last hand-built wall?

Cartesia published part one of its H-Net scaling study: hierarchical networks drop hand-designed BPE entirely and instead learn to chunk raw bytes inside the model through recursive, data-dependent dynamic chunking. The controlled comparison: a 3.67B-parameter two-stage H-Net (Mamba-2 encoder/decoder around a 40-layer Transformer main network, ~7.6x compression) reaches SmolLM3-3B’s final average score at roughly 1T tokens versus 11.1T for the BPE baseline trained to within 2% of the same FLOPs - a 10x-plus efficiency gap. On code the median compute advantage is 2.04x in-distribution and 3.10x across 10 unseen languages (peaking at 5.37x on Prolog); on DNA it transfers zero-shot at 5.95x to 7.06x. 81 points, 4 comments - the classic high-upvote, low-comment deep-research post. If you work on efficient architectures or long context, this “learn the tokenizer too” line is worth reproducing. Source · HN discussion

26. AlpineBSD, a new OS pledging zero AI-generated code

Even operating systems flaunt a no-AI label now?

AlpineBSD combines the NetBSD kernel, libc, and platform tools like mount with an Alpine-style userspace: apk-tools for packages, busybox for the base commands (leaning GNU - GNU make over BSD make), and s6 as init, with OpenRC planned as an alternative. The stated mission is letting users run as little AI-generated software as possible: no LLM output in the base system or main package repo. Status: no installer, only a high-level installation guide, hardware support explicitly worse than Linux, hosted on sourcehut under sircmpwn - Drew DeVault’s account. 234 points, 216 comments, comments nearly catching upvotes; the argument is clearly about the no-AI pledge itself. For developers tired of AI slop, its package-admission process is the part to watch. Source · HN discussion

27. 26,400 live secrets across 7.25M Gitee repos, at a cost of $3,097

Is your key sitting in that pile too?

Security researcher Ben Zimmermann details on MongoDB’s blog how he enumerated Gitee’s public repositories - the platform API has no project listing, so he combined three-letter username prefix search with Elasticsearch aggregation bucketing to escape the 200-result cap - reaching 1.7M usernames and 7,256,531 repos, then bare-cloned and scanned them with 350 concurrent containers on Modal. The yield: 26,400 verified-live credentials - 9,473 Alibaba Cloud (about one per 766 repos), 3,805 YouTube API keys, 3,170 DeepSeek, 1,489 Zhipu, 1,198 OpenAI, 154 Hugging Face. TruffleHog and Kingfisher each caught only about half, with an overlap of just 1,001. Corporate committer domains include huawei.com (32), microsoft.com (22), and bytedance.com (22). Per-repo density puts Gitee at 3.64 per thousand, above GitLab’s 3.11 and Bitbucket’s 2.39. 29 points, 6 comments. If your team uses Gitee, wire TruffleHog —only-verified into CI today. Source · HN discussion

28. Your AES-256 PDF may have a 9-bit key, crackable in 11 ms

11 ms to brute-force, and they called it encryption?

Rick de Jager of V12 shows several PDF implementations deliver far less than their “AES-256” label. 1. MuPDF before 1.28.5 seeded a 48-bit LCG with the epoch in seconds, and because of which state bits feed each output byte, only 9 bits of timestamp influence the key stream - just 512 possible file keys, and an unoptimized Python brute force takes about 11 milliseconds. 2. Foxit Reader 2026.1.3.36551 hashes a 32-bit timestamp together with 32 uninitialized bytes that turn out to be zeros, collapsing the entropy to the timestamp alone; a multithreaded Rust tool tests 376 million timestamps per second, covering every second since 1970 in under five seconds. 3. OpenPDF 3.0.5 and iText 9.7.1 seed ARC4 with a millisecond timestamp plus JVM free memory, with plaintext /U and /O fields available to verify guesses. All flaws were disclosed; MuPDF is patched, while Foxit’s code dates to PDFium’s initial 2014 commit and may live on in other editors. Old documents cannot be re-secured. 113 points, 61 comments. If you distribute encrypted PDFs from these tools, switch generators and re-issue. Source · HN discussion

29. A data center drinks like a city of 50,000, pushback spreads

Who ends up paying AI’s water bill?

Rest of World, adapting material from Paris Marx’s book Hyperscale, lays out the numbers: a hyperscale data center consumes 1 to 5 million gallons of water a day, on par with a city of 10,000 to 50,000 people, and a 2023 UC estimate put training the original ChatGPT at about 185,000 gallons. The cases: Google’s $200M data center in Cerrillos, Chile planned to draw 169 liters per second - the annual drinking water of all 80,000 residents - and after a December 2019 referendum rejected it 49% to 38%, Google returned in February 2022 with an air-cooling proposal; Microsoft hit similar opposition in drought-rationed Quilicura. In Uruguay, a 72-acre site in Canelones would use 2 million gallons a day during Montevideo’s worst drought in 70 years, and a researcher has sued the environment ministry for details. In Brazil, Caucaia hosts a $10B TikTok facility in a region with drought emergencies in 16 of 21 years between 2003 and 2024. 50 points, 41 comments. For AI-infrastructure siting and disclosure work, water is becoming the second permitting gate after electricity. Source · HN discussion

30. $264 of Opus 5.5 rebuilt Calvino’s cities, mouse input never worked

Identity crisis aside, someone still has to ship it?

Konrad Talik used Opus 5.5 to replicate Piotr Migdał’s experiment - one prompt, a three.js visualization of Italo Calvino’s Invisible Cities - and published both bills. Medium effort: $14.98, about 313K output tokens, 53 minutes of API time, 7,746 lines of code. Max effort with ultracode: $264.64, about 5.3M output tokens (3.65M of them thinking), 242 web searches, 16.5 hours of API time, 31,387 lines added - a 17.7x cost ratio across roughly 11 parallel agents, with the session hitting its limit and resuming four days later; and the mouse input never worked, leaving keyboard-only controls. The post concedes there is no real proof of an “Opus 5.5 identity crisis”; the actual argument is the anxiety of watching skills displaced, with Industrial Revolution artisans as the historical parallel and human Go strength rising after AlphaGo as the counterpoint. 20 points, 41 comments, twice as many comments as upvotes. For anyone banking on “AI does my job,” this receipt is a calibration exercise. Source · HN discussion