DeepInfer, the framework DeepSeek says runs its own production services, is open source: 13,900 stars and counting overnight, with the unusual promise that identical input gives identical output. A researcher proves eight certificate authority keys are compromised by signing a note with them. Mistral ships a voice stack in 50 languages. And the curl project, drowning in reports, moves most of its bug triage into the open.

Today: the engine room, the signed confession, the voice models, the Tor exit nodes that stole SSH passwords, curl’s survival plan, and the 50,000 fake statistics ChatGPT keeps citing.

The interesting releases this week are the ones that were never supposed to happen.

🔧 DeepSeek open-sources its engine room

DeepSeek has never shipped the engine it runs its own models on. On October 16 the repo went up quietly; by Tuesday it was the top story on Hacker News. The README’s first line is the whole pitch: DeepInfer is the high-performance inference framework that powers DeepSeek’s large-scale production services. Not a research toy. The thing they actually run.

Inside: the FlashMLA, DeepGEMM, DeepEP and TileLang kernels DeepSeek has been publishing piecemeal, now wired into one stack, plus DSpark speculative decoding with confidence-scheduled verification, sliding-window bounded replay, and a zero-overhead asynchronous batch scheduler. Two features stand out for anyone running evals or agents: batch-invariant inference, so results do not change with batch composition, and deterministic sampling, so identical input tokens plus a fixed seed give identical output. Supported models today: DeepSeek-V4.1-Flash. Supported hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. The honest gap, printed in the README: a KV cache system is required for a complete service and is not included. This is offline inference with the core serving features, not the whole factory.

The repository  ·  HN discussion

✍️ The CA compromise proven with a signature

How do you prove a certificate authority’s private keys are compromised, without a doubt, in public? A researcher identified only as Wayne chose the cleanest possible demonstration: he signed a message with the private key of each of the eight intermediate CAs. The message: "We have a problem here - Wayne." The proof bundle, with certificates and signatures anyone can verify in OpenSSL, is on the report site.

The keys chain to roots trusted by the Mozilla, Microsoft, Apple and Chrome root programs, meaning an attacker holding them could mint valid-looking HTTPS certificates for any website. Wayne reported to the root programs on October 17. Mozilla blocked the certificates in Firefox through OneCRL in under three hours, no browser update needed, and shipped Firefox 158.0.1 plus ESR releases for stragglers the same week. Chrome blocked them from the 17th, Apple devices from the 19th, and eMudhra has revoked the intermediates. The open wound: clients that do not consume revocation data still trust them until updated. The report site ships test URLs: a certificate error from their two curl checks means your client is safe.

The report  ·  Mozilla’s write-up  ·  HN discussion

📞 Mistral answers the phone

Mistral’s launch page opens with a scene instead of a benchmark: Béatrice calls, explains in French that her doctor gave her a referral letter, and asks the agent to book it. Listen, reason, speak, act, in one streaming pipeline. That is the pitch for Voxtral 3, and the demo framing tells you exactly who Mistral is selling to: contact centers and assistants, not transcription hobbyists.

Three pieces. Voxtral Small Transcribe 3 handles speech-to-text across 50+ languages with diarization, end-of-turn detection and context-biasing, where you feed it keyphrases, product names, jargon, to boost recall without fine-tuning. Mistral’s charts claim it beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on FLEURS and diarization benchmarks; those are vendor numbers on vendor-picked cuts. A 4B Mini variant is priced at $0.0035 per minute. Voxtral Mini TTS 3 generates speech in 35+ languages and 50+ dialects, adjusting emotion and pace to context, and a Pro Voice tier adapts a speaker’s style from 30 minutes of audio. The voice agent SDK ties listening, reasoning, tool calling and speaking into one real-time stream. Deployment is the differentiator: Mistral’s EU-hosted API or your own infrastructure, with control over whether audio is retained.

The announcement  ·  HN discussion

🧅 Ten percent of Tor’s exits were stealing SSH passwords

Friday morning, October 2, 09:35. A researcher connecting over Tor hits the warning every developer has clicked past a hundred times: REMOTE HOST IDENTIFICATION HAS CHANGED. IT IS POSSIBLE THAT SOMEONE IS DOING SOMETHING NASTY. This time someone was. His SSH session had been silently redirected to an interception server, and pulling that thread uncovered 333 infected Tor exit nodes, roughly a tenth of the entire network’s ~3,000 exits, running a coordinated campaign.

The mechanics: attackers had planted Ebury 1.8.3, an OpenSSH backdoor and credential stealer, on the servers back in August. They turned compromised machines into exit nodes, port-filtered so their exits mostly carried unencrypted mail traffic, and redirected outbound SSH to 12 interception servers. Fall for it and the attacker takes your password, watches every command you type, then infects your server too. The researcher reported to the Tor Project’s network health team that Friday evening; their reply: "We’ve not seen an attack like that before." By Sunday the relays were flagged BadExit, and by Monday all infected servers were down. The dip is still visible on Tor’s public metrics.

The field report  ·  HN discussion

🚒 curl stops drowning by triaging in public

Daniel Stenberg does not do drama, which makes the language in his latest post land harder: curl’s maintainers feel "strained, overloaded and brought to the brink of what we can muster. We must act to survive." The flood is vulnerability reports, arriving faster than a small team can assess them, with no end in sight.

The policy change is surgical. Starting October 26, reports deemed LOW or MEDIUM severity go to public GitHub issues instead of the private HackerOne tracker. Only HIGH and CRITICAL stay private. Reports sent to HackerOne that turn out to be below HIGH get closed as informative and publicly disclosed there. His number: over 90% of submissions over the last few years were quickly assessed as less than HIGH, meaning the private pipeline was carrying almost entirely public-grade work. Stenberg’s post does not say where the flood comes from; he calls it an avalanche and leaves it there.

Stenberg’s post  ·  Lobsters thread

📊 The 50,000 fake statistics ChatGPT cannot stop citing

Seven months ago Metehan Yesilyurt bought a fresh domain, stateglobe.com, and filled it with 50,000 statistics pages generated in a single day with GPT-4.1. Every number was invented. Total cost, including 15,000 listicles added later: about $50. Then he watched what the machines did with it.

The split is the finding. Google buried the site almost immediately; impressions collapsed to nearly nothing. ChatGPT never got the memo. Checking his Cloudflare crawl data, Yesilyurt counted more than 72,000 requests in the last 30 days from ChatGPT-User alone, the agent ChatGPT uses to fetch pages live while answering questions. These are not training crawls; each one is a real answer being built on fabricated numbers. More than 330 domains now link to the site, including Shopify, Bluehost and DHL, and the made-up statistics have surfaced in an academic paper and a bank’s investment research. All of this is his own reporting on his own experiment, and he works at an AI-search analytics company, so apply the usual discount. The underlying mechanic, live-fetch grounding with no quality gate, is documented in his server logs.

The write-up  ·  HN discussion

tsrs: TypeScript, ported to Rust. Cloudflare, VoidZero and the port’s author are rebuilding the TypeScript compiler in Rust inside Oxc, in collaboration with the TypeScript team. Their benchmarks: up to 8x faster, ~70% less memory. 358 points. Cloudflare blog →

SimCity, but the zoning board says no. Permit Pending makes you an elected mayor: every project faces hearings, NIMBY lawsuits and elections. The worked example of a rezoning dying over 42 months is painfully real. 624 points. permit.petercai.com →

SlopCodeBench: pricier thinking, worse code. The October run finds Opus 5.5 scoring better on medium effort than on xhigh, which costs 2.1-4.2x more. Astra leads at 16.3% strict pass. Author’s benchmark, author’s numbers. humanlayer write-up →

Meta tests a two-link monthly limit. Facebook Pages without a Meta One subscription are being limited to two organic link posts per month in a live test; Sprout Social’s support docs corroborate. A test, not a policy, for now. Sprout Social docs →