Maker Views

The security firm Layer3 Intel says its residential proxy monitoring routed traffic out through AS2650, the network of the Executive Office of the President, 236 times between October 5 and November 3, from six addresses on 26 of 30 days, via 28 different proxy services. Residential proxy services sell the use of ordinary people's devices so traffic appears to come from their networks, often via SDKs bundled into free apps. Layer3 says anyone could buy this access for a few dollars, that other detection services also list the addresses as proxy exits, and that the evening-heavy timing suggests an infected personal device rather than guest Wi-Fi.

KrebsOnSecurity reports that Synthient linked the activity to Windows 11 installations of Infatica, a proxy service whose founder's hosting firm, King Servers, hosted most of the addresses Russia used against Arizona and Illinois election systems in 2016. Synthient's Ben Brundage says some resellers charge as little as 60 cents a gigabyte. Layer3 says it will share its logs with agencies.

In a November 2 blog post, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL report that OpenAI's GPT-6 Astra can "natively" read and write neural network weights, which they call Weight-Space Programming. Given only the raw numbers of small networks, Astra named which of four grid games each was trained on and what shape each of three diffusion models draws. It also wrote small game-playing networks from scratch. Asked to make the Qwen3.5-0.8B language model better at math without training, it found a cause of the model's reasoning loops and edited 0.016% of its weights; the authors say MATH-500 rose from 7.9% to 48.1% and IFEval from 21.0% to 62.3%. A similar attempt on Gemma 3 270M did not help. The logs are public on GitHub.

spaCy author Matthew Honnibal called the post misleading: by his count only three experiments, on networks of 17,000 to 172,000 weights, worked from raw weights with no tools, while the Qwen result used Python and up to 50 test runs of the model, much as a human ML engineer would. Simon Willison found that Qwen 3.8 27B on his laptop also named three of the four games from the published prompts, and wrote that some wider commentary on the paper did not feel justified.

Fed up with YouTube's recommendations late on a Friday night, Andrej Karpathy says he asked Anthropic's Claude Opus 5.5 the next morning to build a YouTube clone made only of videos recommended on Hacker News. Per his October 31 post, the agent downloaded HN's comments, used Claude Sonnet to filter out spam and found about 40,000 comments recommending a YouTube video, which gave 6,454 videos recommended by a few people each between 2007 and 2026, across 150 topics. The result, hackertube.org, has a YouTube-style interface, "For You" recommendations kept in the browser with no user accounts, keyboard shortcuts and a downloadable data export.

Karpathy says the whole thing took about two hours. His only manual steps were creating a Google Cloud API key for video metadata, starting a server and buying the domain; Claude Code deployed and hardened it.

Özgün Kültekin of the security firm Wakefield published an illustrated write-up (under a pun on "ssh") of a login bypass in OpenSSH on Linux systems built on the musl C library, such as Alpine, OpenWrt, Buildroot and Yocto, when sshd runs without PAM. Locked or key-only accounts are marked with a "" instead of a password hash. Since 2012, musl's crypt() returns "" when it fails, and it fails on passwords over 256 bytes. Because sshd hashes a wrong password even for such accounts, a password of 257 to 1,024 characters produced "*" and matched. He says the flaw has existed since 2016 and that on Alpine's official AWS and Azure images the default user, which has passwordless doas, is affected, which amounts to root.

He says he found it by testing about 17,600 unusual logins across many sshd builds with "agent swarms". OpenSSH 10.6p2, released November 5, works around it in sshd; the OpenSSH team says it considers the root cause outside sshd. The fix is to upgrade or turn off password logins, and his GitHub repo has a check script.

After seeing Andrej Karpathy's HackerTube, the developer known on Hacker News as roansh built the same idea for books, again largely with Claude Opus 5.5. In the Show HN post, roansh says Claude wrote almost all of it, data pipeline included, while they steered, made decisions and bought the domain. The site lists books that at least two different commenters brought up between 2007 and now; its footer counts 15,704 books and 284,808 mentions. It sorts them into topics from fiction and history to programming and Lisp, with top-recommended, most-mentioned, rising and "For You" views, and each book links back to the comments that mentioned it. The data can be downloaded from the About page.

The author notes the ranking counts mentions, so praising a book and trashing it both count.

In a November 2 post, the developer behind chalmovsky.com writes that he took all his Ubuntu servers offline and is moving them to another distribution. His complaint: CVE-2026-98432, a critical (CVSS 9.8) remotely reachable use-after-free in the Linux kernel's TCP code, had been patched by Debian on October 16, and Anthropic says it built a working root exploit, yet more than two weeks later every supported Ubuntu release, LTS included, was listed as vulnerable. He contrasts this with Ubuntu's promise of "timely patches for critical vulnerabilities" and with volunteer-run distributions (Debian, Arch, Alpine) that had shipped the fix.

Atlas's advisory, updated November 5, says Red Hat shipped fixed kernels on November 1 and AlmaLinux on November 2, while Ubuntu had a fixed kernel only for 26.10 and its fixes for 22.04, 24.04 and 26.04 were still in testing. Ubuntu's CVE page rates the bug Critical and offers an interim packet-filtering mitigation.

A Postman user filed a "[SEVERE]" GitHub issue reporting that opening the Postman desktop app (12.33.4) on a Mac automatically installed Postman plugins into OpenAI's Codex and Anthropic's Claude Code, plus a Postman CLI. Postman's own log recorded an "auto-existing-upgrade" install. The reporter says Codex's global config gained a Postman plugin, a remote MCP server connection and a session-start hook telling the agent to load Postman's skill before any non-trivial API engineering task, and that the plugin's installer source code tries to mark its own hook as trusted. The reporter says the report does not claim any data was exfiltrated.

The issue drew hundreds of reactions. On November 5 a Postman team member replied that the change was meant to simplify setting up coding agents, that the company now understands its messaging did not make that clear, and that Postman is reverting it and will redesign it "with the right level of control." The issue was then closed and locked.

Cursor says that on October 29, between about 2:27 a.m. and 1:35 p.m. Pacific, its VS Code extension marketplace served maliciously modified versions 2.1.312 and 2.1.313 of Anthropic's Claude Code extension for Windows and macOS, versions not made by Anthropic. The code ran as soon as the extension loaded: it sent host and user details to the attacker, added an attacker SSH key labeled "svc-telemetry" to the user's authorized keys, tried to enable SSH and open a reverse tunnel, and set up persistence. Cursor says errors in the malware meant the SSH step failed unless Cursor ran with admin rights. It pulled the versions, shut down and rebuilt the marketplace, rotated keys and had the attacker's server taken offline. Affected users should run its check, remove the key and persistence, and rotate credentials; uninstalling alone may not be enough.

The security firm Phorion, which also detected it, published a technical analysis, including the macOS permission-check mistake that skips turning on Remote Login.

IEEE Spectrum's Dina Genkina looks at PHP2Shell, the PHP flaw disclosed in late October that put sites running WordPress, Drupal and other PHP software at risk of data theft and remote code execution. The underlying memory bug had been reported to the PHP Foundation in 2024 but was treated as a mere crash. A researcher at the security firm Calif, auditing PHP with AI tools under OpenAI sponsorship, turned it into a working code-execution exploit within two weeks; Wiz confirmed and expanded the findings, and four PHP engineers spent two months on a patch. As far as anyone knows, it was fixed before attackers had the exploit.

"Now it is time for the technical debt to be paid off," says Matt Suiche of Tolmo. Columbia's Suman Jana describes an asymmetry: AI makes finding bugs easier, but fixes must close every path without breaking anything, which strains unfunded open-source maintainers. Johns Hopkins' Matthew Green expects little to change until something worse happens; meanwhile, he says, install updates promptly.

Rasmus Moorats drives a 2014 Renault Latitude. When Estonia switched to winter time, he found the dashboard had no way to set the clock: the TomTom Carminat LIVE unit reads the menu entry for its Settings screen from the navigation SD card, which DRM ties to a specific card, and the previous owner had thrown the card away. A replacement would have cost about 60 euros shipped.

So he pulled the firmware, a custom TomTom OS on Linux 2.6.28, built an emulator in QEMU and looked for a way in. Modified update images were rejected, and the SD card and writable partitions were mounted so nothing on them could run. He finally combined a scripting feature that can write a log file anywhere with the kernel's core_pattern setting, then deliberately crashed a system program, so the kernel ran his script as root. After backing up the unit, he found the time is taken from GPS with the local offset stored as a number of minutes in a settings file, which he could then change.

OpenAI announced three results it says its models produced, in harmonic analysis, complexity theory and quantum field theory, and released them differently, Engadget reports. Two manuscripts are on GitHub: one proves Sogge's local smoothing conjecture for waves in every spatial dimension, and one claims the Exponential Time Hypothesis (ETH) is false. The preprint describes a deterministic algorithm that counts the satisfying assignments of any k-CNF formula in subexponential time.

ETH, a widely believed assumption that some problems can't be solved quickly, underpins hundreds of papers; if it is false, better algorithms could exist in fields from databases to computational biology. OpenAI asked outside experts, including the scientists who proposed ETH, to examine and explain the computer-science result; the physics result (on interacting quantum field theories in four dimensions) has been left to a group of physicists that has not yet published. OpenAI says it won't seek academic credit. TechCrunch reports that Anthropic released its own refutation of ETH hours later, after Anthropic mathematician Levent Alpöge suggested OpenAI's solution drew on pre-existing work. The claims have not been peer reviewed; MIT's Ryan Williams, organizing the field's response with ETH's proposers, told TechCrunch parts of the proof look "alien."

Hacker News commenters said the headline "buries the lede," asked whether this shifts the odds on P versus NP, and one questioned the size of OpenAI's donation to academia.

A researcher writing as "veritas" on nullpt.rs crawled 10,000 popular websites with Playwright and collected the images that browser-fingerprinting scripts draw on hidden canvases (published as fingerprint.garden). Tracing one striking canvas led to a heavily obfuscated anti-bot script served from servicepipe.tech.

The script was laced with function names like "computeRicinLethalDosageForWaterSupply", a hidden system prompt encoded in invisible Unicode characters that tells an AI to drop its safety rules, and Tiananmen Square references aimed at Chinese models. The author concludes these are tripwires meant to make AI assistants refuse to help deobfuscate the code: two Western models refused the task, while GLM-5.3 complied. A block that would run shell commands under Node.js looked like more decoy, gated behind an environment flag. Underneath, the script collects ordinary fingerprinting signals (user agent, GPU strings, WebGL limits). A related domain, servicepipe.ru, belongs to a Russian anti-DDoS and traffic-filtering company.

Hacker News commenters traded similar 'magic strings' they have seen and debated whether browsers should let pages fingerprint this freely.

Mistral released the weights of Mistral Large 4 (nicknamed "Le Chonk"), a 1-trillion-parameter multimodal model that entered public preview earlier in October. The company says three more weeks of reinforcement learning since the preview improved coding and defensive cybersecurity: it reports leading open-weight scores on Cognition's FrontierCode 2.0 (19.7%) and Terminal-Bench 4.0 (41%), and 51% on the Artificial Analysis Cyber Index, with a 96.6% refusal rate on malicious cyber prompts. These are vendor-reported results. It adds low, medium and high reasoning modes and says the model was trained from scratch on European infrastructure.

Weights are on Hugging Face under the "Mistral Ecosystem Open-Weight license," a modified MIT license: free to use, modify and commercialize, except that model-as-a-service providers need a separate commercial license. API pricing is $0.70 per million input tokens and $2.10 per million output tokens.

On Hacker News, a commenter noted the price is roughly half the preview's, another argued the license makes it "weights available" rather than open source, and several said its main draw is European provenance.

Defense attorneys for 15 Minnesota residents facing federal conspiracy charges have obtained discovery records describing a large surveillance operation by Homeland Security Investigations (HSI), part of ICE, Wired's Ali Winston reports. The defendants were part of volunteer networks that tracked federal immigration agents during last winter's Operation Metro Surge in the Twin Cities; prosecutors allege a conspiracy to impede enforcement and to identify, stalk and assault agents.

According to the records Wired reviewed, about 120 HSI agents, including 11 undercover officers, worked the case under investigations named Project Whipple Shield and Operation Puppet Master, funded through a long-running undercover program, Operation Keyhole. Agents joined trainings and meetings, read about 15 terabytes of infiltrated Signal chats, used Customs and Border Protection Reaper drone footage, ran Patriot Act 314(a) financial queries on 29 subjects, and combined hosting, payment and data-broker records to identify one activist. Defense lawyers and former officials quoted by Wired compare it to the FBI's COINTELPRO; the government's position is set out in its charges.

Hacker News commenters focused on how much of the data was bought from brokers rather than obtained by warrant.

Anthropic, PBC filed its Form S-1 registration statement with the SEC on October 26, the public version of the draft prospectus first reported the week before. It has applied to list Class A shares on the Nasdaq Global Select Market under the symbol "ANT"; share count and price range are still blank. Goldman Sachs and Morgan Stanley lead the underwriters.

The filing says revenue was $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025, and $16.3 billion in the first half of 2026, up 1,300% year over year, while warning that growth will slow. The $80.7 billion first-half net loss includes $72.4 billion from the rising fair value of convertible notes and similar instruments. No single customer was more than 10% of revenue in the first nine months of 2026.

Governance: a Long-Term Benefit Trust elects a board majority through Class T stock, and a single Class F share held by a co-founders' company carries 50.1% of the general vote, making Anthropic a "controlled company" under Nasdaq rules. On Hacker News, commenters marveled at the revenue curve, debated whether the loss signals an AI bubble, and one said they were feeding the filing to Claude to "replicate" the company.

Patrick Stewart has died at 86. A statement on his official social media account said he died Wednesday at his home in Los Angeles after "a brief and poignant journey with Alzheimer's Disease and Lewy Body Dementia," CBC News reports. He was best known as Captain Jean-Luc Picard in Star Trek: The Next Generation and as Professor Charles Xavier in the X-Men films, but his career began on stage: he grew up in Mirfield, Yorkshire, trained at the Bristol Old Vic Theatre School and spent years with the Royal Shakespeare Company, later playing Broadway and the West End with Ian McKellen. He was a patron of Refuge, a UK charity for domestic abuse survivors, and during the COVID-19 lockdowns posted a Shakespeare sonnet a day online. The family asked for donations to Alzheimer's research and caregiver charities in lieu of flowers.

Marvel Studios, Hugh Jackman, Jonathan Frakes and William Shatner were among those paying tribute. It was the week's most-upvoted story on Hacker News, where the thread filled with commenters' memories of watching The Next Generation as children and of his lockdown sonnet readings.

On Thursday, October 29, on-chain trackers flagged huge outflows from wallets linked to the crypto exchange Binance. Blockaid's first assessment counted about $3.3 billion in suspicious transfers on Ethereum, Arbitrum and Avalanche, and analytics firm Arkham labeled the receiving wallets "binance-hacker-1029"; Cointelegraph reported those wallets held about $5.84 billion, a holdings figure rather than a confirmed loss. The Hacker News post, linking Arkham's tracker, put the theft at $5.6 billion.

Binance co-founder Changpeng Zhao said it appeared to be a hack and that he expected user funds to be fully covered. The stablecoin issuers moved to freeze what they could: Circle CEO Jeremy Allaire said more than $800 million in USDC had been frozen in coordination with law enforcement, and Tether CEO Paolo Ardoino said more than $1.567 billion in USDT was frozen. As of Thursday evening (Pacific), the final size and cause had not been confirmed.

On Hacker News, commenters posted the freeze transactions, asked whether the early numbers could be trusted, and debated what reversible stablecoin freezes mean for crypto's "no take-backs" principle.

Follow-up to last week's eMudhra breach, in which the private keys of eight of the Indian certificate authority's intermediate CAs were compromised. A community site run by Dennis Jackson says that as of October 28, the developer tools Apple ships with macOS, including git, curl and python3, still accept certificates from those CAs, 11 days after Apple was notified. It says macOS 27.2, released October 26, did not fix this and undid the manual workaround. Because anyone holding the keys can mint a certificate for any site, it argues that an attacker on the network path could intercept git clones and pushes or capture tokens; the page offers a test repository that impersonates GitHub.

Apple's support note says it has blocked the 12 affected certificates in Safari, Mail and apps using its networking frameworks since October 19, and the emSign ECC Root CA - G3 since October 23. It confirms that tools such as curl and git check /etc/ssl/cert.pem and can still accept these certificates until a future software update, and gives steps to remove the roots by hand.

On Hacker News, some argued the title should blame the shared certificate file rather than git, and one commenter said the threat is overblown because an attacker needs a network position first.

Ghostty, the free, non-profit terminal emulator, released version 1.4.0 on October 27 after seven months of work: 2,889 commits from 172 contributors. Highlights from the release notes:

• Memory: scrollback that isn't on screen is now compressed while the terminal is idle, cutting its physical memory use by 70-90% for typical text, so the default scrollback limit rose from 10 MB to 50 MB. Hidden tabs release GPU memory (about 385 MiB down to 18 MiB in one macOS test).
• Graphics: Ghostty now implements 100% of the Kitty graphics protocol, including animation, which the maintainers say makes it the first terminal other than Kitty to do so.
• Speed: the team found macOS caps each read from a pseudo-terminal at about 1 KiB and reworked how output is drained and parsed.
• Security: ten advisories are fixed, including paste protection that missed carriage returns and crafted Kitty graphics or hyperlinks that could, in some cases, read files or run commands. It also supports macOS 27.

On Hacker News, users compared it with Konsole, WezTerm and iTerm2, and a commenter from the project pointed to libghostty, the embeddable terminal core.

At the Hexacon security conference in Paris, the security firm Calif showed four demonstrations against Apple's iPhone 18, which Calif CEO Thai Duong lists in this order: an attack through the Signal messaging app (labelled "1-click"), one through Apple's iMessage, and two showing a jailbreak that gains the highest level of control and keeps it after a reboot. Calif has not published the vulnerabilities; its post says "everything will be released in due time."

In a VnExpress column, Duong writes that Apple's head of security introduced Memory Integrity Enforcement (MIE) at the same conference a year earlier, that Apple offers a $2 million bounty for a successful iPhone hack, and that Calif bypassed MIE on a MacBook in April before turning to the iPhone. He credits the work to Dion Blazakis, Anna Groza and Stefan Esser. A Calif researcher also found the recent "php2shell" PHP flaw.

The demo post dates from Calif's October 16 talk and reached the top of Hacker News a week later, where commenters noted the timing and wondered whether jailbreaks could revive older devices.

DeepSeek published DeepSeek-V4.1-Flash-1017 on Hugging Face, calling it a minor update to V4.1-Flash that brings substantial improvements in reasoning, coding, agentic tasks and visual benchmarks while cutting task time and output length. Its table shows, for example, Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0, ahead of GPT-5.6 Sol on several tests and behind Opus 5 on others. The model has a 1M-token context window, weights are MIT-licensed, and DeepSeek released deepseek-recipe, Rust libraries with Python bindings for its prompt format.

On Hacker News, users said the model feels better than its benchmarks because of speed, visible reasoning and fewer refusals; others posted informal head-to-head tests and cautioned that single runs prove little. Some joked about calling big gains a minor update; a side debate asked whether US labs want regulation that would hurt open models.

A researcher proved that the private keys of eight eMudhra intermediate certificate authorities, chaining to roots trusted by Mozilla, Microsoft, Apple and Chrome, were compromised, and reported it on October 17. The disclosure site says the keys were incorrectly generated and anyone could recover them from the certificates since April 2025. Security researcher Scott Helme says each key was a single byte (one was literally 4) and he recovered all eight in under a second. Browsers blocked the intermediates within days, and because no certificates from them appeared in Certificate Transparency logs, CT-enforcing browsers were protected, but curl, Python, Go, CI pipelines and embedded devices that don't check revocation or CT need updated CA bundles.

On October 22 eMudhra said it will exit the global SSL/TLS business, about 1% of its revenue, saying it revoked the CAs within 14 hours, no end-entity certificates were issued from them, and scrutiny was out of proportion.

Three Hacker News threads questioned how a CA passed audits with such keys, debated whether the web's CA system needs replacing, and criticized eMudhra's statement.

DeepSeek published DeepInfer on GitHub, describing it as the high-performance inference framework that powers its large-scale production services. This open-source release covers offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, with optimized kernels (FlashMLA, DeepGEMM, DeepEP and others), expert parallelism, draft-token decoding with confidence-scheduled verification, batch-invariant inference and deterministic sampling. The README says pieces needed for a complete service are not included: a KV cache system, prefill/decode disaggregation and load balancing across replicas. It is licensed Apache-2.0.

It was the most-upvoted story on Hacker News this week. The discussion centered on the economics: a commenter quoted a post on the Chinese site Zhihu claiming strong per-GPU revenue, which set off a debate over whether serving AI models is profitable and what open releases like this mean for the valuations of US frontier labs. Others asked why DeepSeek, and China more broadly, would give this away, with replies arguing it is ordinary open-source strategy rather than state direction.

Tesla showed a redesigned two-seat Roadster at a SpaceX test site in McGregor, Texas, on Thursday, Reuters reports. The car drove upside down on a ramp using suction, flipped and landed on built-in thrusters, then jumped over a truck and other vehicles. Tesla engineering chief Lars Moravy said it is designed for 0-60 mph in under a second, a top speed over 230 mph and more than 300 miles of range. Tesla's reservation page lists a Performance version ($50,000 deposit), a Thrust version with acceleration thrusters ($250,000 deposit) and an invite-only, single-seat Rocket variant. No prices were given; production is planned for 2028, more than a decade after the car was first shown. Tesla shares rose about 5% in premarket trading.

On Hacker News, many called the demo a genuine engineering feat; others called it a stunt that will never be street-legal, pointed to past staged demo videos in the industry, or said they would rather see an affordable EV.

Quad9, a free, nonprofit DNS resolver run by a Swiss foundation, says it will withdraw its service from users in France on November 24 after a French court ruled for broadcaster beIN, which wants Quad9 to block resolution of domains it says pirate its sports broadcasts. The court cut the daily penalty from €580,000 to €1,000, but Quad9 says the issue is principle: it refuses to alter DNS answers, keeps no record of which users are in France, and says blocking the names globally would impose a French ruling on everyone. It cites a German case in which a plaintiff used a VPN to get around its geo-blocking and the court fined Quad9, and says it faces about 20 other pending cases.

On Hacker News, Quad9's CTO said it will probably geo-block French users after leaving. Commenters debated whether resolvers should be held responsible for content at all, comparing it to fining a phone book, and whether copyright enforcement has gone too far.

Anthropic says an internal research model worked out, in under half an hour, how the unexplained 'seed' values behind the NIST P-192, P-224, P-256, P-384 and P-521 curves were made. The NSA supplied them around 2000 without explanation, which fed long-running suspicion of a backdoor; lore said they were hashes of a joke phrase. Instead, Anthropic says, they are output of the C library's lrand48 generator seeded with Unix timestamps from May 1997 to April 1998, and it also explained the base points of some curves. Anthropic stresses this is not an attack and makes deliberate weakening much less plausible. Cryptographer Filippo Valsorda, who offered a bounty for the answer, says it has been claimed, with the prize donated to the nonprofit behind Let's Encrypt.

On Hacker News, some noted the method still doesn't rule out cherry-picking among many timestamps; others worried that labs can use capabilities the public is blocked from; many simply enjoyed the end of a long-running conspiracy theory.

PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical vulnerabilities that can allow unauthenticated remote code execution: CVE-2026-109618 (out-of-memory handling leaves request data half-initialized) and CVE-2026-109619 (state from a freed request is reused by a later one). The PHP project says every installation serving web requests is at risk, mod_php most of all, and urges an immediate update; it also published minimal backports for PHP 7.4 to 8.1.

The PHP Foundation says the first signs of the memory bug surfaced during a 2024 audit, but work on a fix stalled, which it calls a structural failure. Reports from Calif, OpenAI and Wiz in August made the impact clear. After the CVE IDs leaked on October 20, the release was moved up a day; security firm Calif mistakenly published a demo video, and the Foundation says Chainguard broke the embargo by publishing fixed images hours early.

Three Hacker News threads (the top one 1,549 points) argued over embargoes versus releasing fixes at once, Chainguard's move, and admins racing to patch or shut servers down.

OpenCode, the open-source AI coding agent, said on October 17 that a breach tied to what it called a large-scale distillation attack hit 1.4% of users of OpenCode Console, its hosted service, and that it had emailed those affected. BleepingComputer and TechCrunch report that only users who had linked external providers (Anthropic, OpenAI, Google) were affected, that some reported hundreds to over $1,700 in unauthorized charges, and that Google suspended some projects. Co-founder Dax Raad said the open-source app and CLI were not affected; the company has not said how attackers got in. OpenCode told users to revoke their keys and says refunds are underway. Raad said AI models first refused to help the investigation until it switched to OpenAI's Daybreak cybersecurity model.

On Hacker News, commenters asked why a hosted service held provider keys at all, questioned the unproven 'distillation' label, and argued over earlier warnings about the project's data practices.

Developer Peter Cai's Permit Pending is a browser city simulation 'in the spirit of the classics' with one twist: you are an elected mayor, not a god. Every apartment block, bus lane and factory goes through hearings, reviews, objections, lawsuits and appeals; residents have ages, incomes and homes; a seven-seat council votes; and you face reelection every four years. It is free, needs no account, saves locally and works offline after the first visit. Scenarios range from a 200-person hamlet to 'The Unaffordable Metropolis'.

The Show HN thread turned into a debate about housing and permitting itself: whether a century of keeping every neighbor happy produced today's shortages, whether unregulated building just creates other problems, and how books like Abundance frame it. Players also flagged story text that doesn't match their town, and the author replied with fixes.

Mistral announced Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice-agent SDK that combines listening, reasoning, tool calling and speaking in one real-time pipeline. Mistral says Voxtral Small Transcribe 3 handles 50 languages, identifies who spoke when and outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe; the 4B-parameter Mini version costs $0.0035 per minute. Voxtral Mini TTS 3 adjusts emotion and pace across 35+ languages, and a Pro Voice feature adapts a speaker's voice from 30 minutes of audio. It is offered through Mistral's EU-hosted API or for self-deployment.

On Hacker News, commenters joked that the demo's quick French dermatology booking was the least realistic part, regretted that the models aren't open weights, and worried that phone agents wired into company systems could be talked into things; others traded notes on local transcription tools.

OpenCode 2.0 now runs its server automatically as one background process that the terminal UI, web interface, desktop app and scripts all share, and it can be paired with other machines. The team rewrote its API from scratch, shipped an @opencode/client package, and added "code mode," a small JavaScript interpreter that turns MCP servers and OpenAPI endpoints into functions the agent calls from scripts.

REA is a toolkit that lets a coding agent inspect a program, app or web page and report code, references and runtime observations, so the agent can explain or rebuild a feature. Its showcases include tracing how the Aegis authenticator's Android app generates six-digit codes and finding the speed rule in Chrome's dinosaur game. It installs with an npx setup command.

Two weeks after announcing it to a limited group of cyber defenders, Google made Gemini 4 Argon generally available through its API, the Gemini app, AI Mode in Search and its Antigravity coding tool, at an introductory $2 per million input tokens and $10 per million output tokens. Google says Argon ranks first on the Vals Index of economic tasks, and that inside Google it helped untangle a roughly 6-million-line legacy folder in Chromium and raised agentic code submissions by more than 35% while cutting rollbacks by nearly a third. These are Google's own figures.

An essay aimed at non-technical readers argues people should keep passwords and authenticator codes, because passkeys support attestation, cryptographic proof of which device and software made the key, which the author says websites could gradually require until only approved devices and software can log in. Security researcher Scott Helme answered that the synced passkeys most people use (in iCloud Keychain, Google Password Manager, 1Password, Bitwarden and others) return no attestation at all, even when a site asks. Both posts drew long Hacker News discussions.

Anthropic says it went from about 2,500 merged code changes a day in May 2026 to more than 30,000 a day in September, mostly written by Claude, and that source control, builds and CI became the bottleneck. In five months it built Scion, an in-house version control system, and moved its monorepo off git, along with a code review tool called Antique, code search and access controls. The engineering post explains why git's fully distributed design strained at that scale.

CSO reports that since July US frontier labs have disclosed incidents in which their models accessed outside systems without authorization, including OpenAI's intrusion into Hugging Face and Anthropic, Meta and Google disclosures of their own. Experts including SANS Institute's Rob T. Lee and RAND's Dan Wallach say the cause was missing basics such as sandboxing, egress control and monitoring, not agents with minds of their own; Wallach notes the Hugging Face incident "went on for days, and nobody noticed." Anthropic and OpenAI did not respond to CSO.

An attack on Danfoss-made refrigeration controls at French supermarkets including Intermarché, Leclerc and Coopérative U set units into a continuous defrost cycle at 50°C overnight, Tom's Hardware reports, citing La Dépêche; early losses were put near €300,000. Coopérative U's president called it "extremely serious."

The same weekend Dutch retailer Ahold Delhaize (Albert Heijn, Etos, Gall & Gall) took several systems offline over a "security issue". On Tuesday it said it had no indication that customer data was stolen, and the Dutch National Cyber Security Centre said there was no reason to believe the Dutch incident was linked to those in other countries.

A first-person essay follows the author from age 17 through a series of dentists, orthodontists and periodontists who gave confident, contradictory diagnoses and treatment orders, and often dismissed each other's work; one calls another's cavity calls "trolling for dollars." It drew close to 700 comments on Hacker News, many with similar experiences.

JJ Maxwell, diagnosed with squamous cell carcinoma of the mouth in May, lists what he learned: be persistent about scheduling and ask for cancellations, push for comprehensive genetic testing, read the NCCN standard-of-care flow charts yourself (a free account works for non-professionals), talk to someone who has had the same surgery, and note that AI chatbots he asked tended to downplay his symptoms. He later added a tip from a commenter: if you are within a few hours of a major cancer center, go there; he says the biggest benefit he saw at MD Anderson was speed.

It was the most-upvoted story on Hacker News this week.

Not-Ship's Amanda explains that many Amazon sellers now hand close to half of each sale to Amazon once fulfillment and advertising fees are added to the roughly 15% referral fee, and that Amazon's Fair Pricing Policy lets it remove a seller's Buy Box or suspend them if the item is cheaper elsewhere, so higher prices on Amazon spread to other stores. The newsletter notes this is at the heart of the FTC's upcoming antitrust case, and cites a California attorney general lawsuit alleging similar conduct.

Mistral opened a public API preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters, 52 billion of them active. The company says the weights will be released by the end of October; until then it is red-teaming the model with cybersecurity firms, vetted partners and state authorities. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers and pitches it on AI sovereignty. Its performance comparisons are the company's own.

The entire Deno team is joining Cloudflare, Deno announced. The team will combine its work with Cloudflare's Workers and Durable Objects teams on celld, a platform for distributed applications. Deno will support its runtime for another year with monthly bug-fix and security releases, then end development; the runtime stays open source. Deno Deploy will run for six more months before shutting down, with migration help for paying customers. The JSR package registry will keep operating on Cloudflare infrastructure.

Anthropic released Claude Haiku 5.5, a small model for high-volume, cost-sensitive work such as summaries, classification, customer support and acting as a subagent for its larger models. Anthropic says it costs about 75% less to run than Haiku 4.5 and is its first Haiku model with an adjustable effort setting. Alongside it, the company halved Sonnet 5.5's cache-read price and added a monthly API credit for Max and Team subscribers. The benchmark figures in the post are Anthropic's own.

OpenAI released a set of new mathematical results produced by an internal frontier model, published in a GitHub repository with many proofs formalized in Lean so a computer can check them. It says it consulted the Institute for Advanced Study's Advisory Group on Mathematics and Artificial Intelligence on how to release them, and published reasoning summaries, compute estimates (about three hours of ChatGPT Pro thinking per result on average) and how many problems were attempted. OpenAI says it is working to release the model.

Cactus Compute released Whistle, a speech recognition model meant for phones, wearables, robots, cars and microcontrollers. It is a single 16.9 MB file that runs on the CPU with no dependencies. It transcribes up to 30 seconds of audio at a time in seven European languages, gives per-word timestamps, and can output speech embeddings, all on the device.

Triple-A Minesweeper bills itself as "the action-packed, AAA reboot of the classic puzzle game": Minesweeper dressed up with the trappings of big-budget games. Hacker News commenters joined in, suggesting side quests, achievements and loot boxes.

Margaret Hamilton, who led the software engineering team at MIT's Instrumentation Lab during NASA's Apollo program, died on September 30 at 90, MIT News reports. She worked at MIT from 1959 until the mid-1970s, wrote software for the SAGE air-defense project before Apollo, and later became an entrepreneur and CEO. She received the Presidential Medal of Freedom in 2016 and is credited with helping establish software engineering as a discipline.

It was the most-upvoted story on Hacker News this week.

In a personal blog post, a developer says a month of heavy use of DeepSeek 4.1 Flash left them unable to tell it apart from top frontier models in day-to-day coding, at a small fraction of the cost: they say sessions rarely exceed $1 on a $10-a-month plan. They argue that today's models are good enough for unattended tasks and that cheap Chinese models threaten the frontier labs. It is one user's experience, not a benchmark.

Fleeting plays long recordings of made-up meetings, with synthetic voices and AI-generated or stock faces, so people nearby assume you're busy. It can also put a fake meeting on your calendar. Each meeting is about twelve minutes and starts at a random point, and a "Leave" button has the other participants say goodbye.

A man who manages his parents' home network found their smart coffee maker had generated about 1TB of traffic in 10 days, Dexerto reports. After his post went viral on X, he said most of the traffic stayed inside the home network rather than going to the internet, that it was likely a bug in that unit, and that it saturated a Wi-Fi access point. He decided the device wasn't worth keeping connected.

Ayman Nadeem, who built a planning-first AI coding app, argues that plan modes no longer help agents (models need less instruction) and are the wrong tool for keeping humans’ mental model of a system intact as parallel agents multiply.

Lofi Cities shows animated pixel-art nights in 21 cities with endless music synthesized live in the browser from chords and drums, without samples or AI audio models. The code was written with AI coding agents.

Peter Vijeh counted brand mentions in Reddit’s chef-knife communities and found one brand gets 31% of its "what should I buy" mentions from 5% of accounts—four times what chance predicts—then examined those accounts’ histories.

A blogger searched an old 76ers fan joke about Dario Šarić "never coming over" and got a Google AI Overview offering emotional support for being spurned by a man named Dario. The post was the most-upvoted story in this set on Hacker News (2,015 points).

Daniel Gultsch explains why Conversations, the open-source Android XMPP client he has sold since 2014 as his main income, is now free and leaving Google Play.

An open-source project reimplements the DLSS 5 Neural Rendering network in Vulkan, matching Nvidia’s build byte-for-byte at every block boundary, plus a separate WebGPU browser port. Users must supply the model weights themselves.