The day's tech, sifted: Oct 30, 2026

What matters today: Binance says a hack drained about $6.3B from its wallets, possibly the largest crypto theft ever, and TRM Labs attributes it to North Korea's TraderTraitor; Binance pledges to cover every loss. The AI-offense story is the same shape: CrowdStrike says China-linked hackers used an AI tool to target 200+ organizations in 11 days, and Epoch shows an open-weight model turning unshipped Firefox fixes into working exploits.

Security & Privacy

• Binance hack: Binance suspended all deposits and withdrawals while Tether and Circle froze billions in coins, then said the theft was about $6.3B. TRM Labs says the funds moved across 10 blockchains, mostly in 20 minutes, beating February 2025's $1.5B Bybit hack; 808bits reports the hot wallet signed 50 EIP-7702 delegations (a feature that lets a wallet hand execution rights to a contract) three minutes before the drain. Because Binance is covering losses, the damage lands on its balance sheet and on bitcoin and ether prices, which fell.
• ARTEX espionage campaign: CrowdStrike says the China-linked group VAPOR PANDA used an AI tool called ARTEX, driven by a DeepSeek model and another LLM, to target 200+ organizations in 11 days and compromise at least nine; the technical write-up frames it as AI scaling a global campaign, so reconnaissance that took weeks now fits in days.
• Patch gap: Epoch reports GLM 5.3, an open-weight model, built Firefox exploits from published but not yet shipped fixes across six releases for about $1,000 each. The window between a public fix and users receiving it is now cheap to attack.
• Polymarket incident: Polymarket resumed trading after a breach at a third-party provider, which a source names as DebugBear, and says it will reimburse affected traders.
• Trojanized Claude Code extension: A fake Claude Code extension plants a persistent SSH backdoor; only install editor extensions from the verified publisher.
• Pentagon agent access: Wired reports a defense contractor gave a personal AI agent access to a secure Pentagon system, the insider-risk version of the agent-permissions problem.

AI / LLMs

• Gemini 4 benchmarks: Cognition's FrontierCode 2.0 (a test of whether code would actually be merged) has Gemini 4 Carbon at 40% against a best score of 45%, and Artificial Analysis's new HLE-Diamond board has Carbon on top at 69% with open-weight models more than 30 points behind. Frontier models cluster near 70% there, so the gap that matters is open versus closed.
• ChatGPT credits for the API: OpenAI now gives its $100, $200 and $500 subscribers matching API credits, copying Anthropic's earlier Max and Team benefit, so a subscription doubles as prototyping budget on either side.
• Token pricing: The New Stack reports Sam Altman nearly scrapped token pricing before backing off, so per-token billing survives for now.
• Open training stack: The Institute of Foundation Models released K2 Horizon 360 under Apache 2.0: 380 checkpoints, 13T training tokens and RL infrastructure, so outsiders can reproduce the whole recipe, not just run the model.
• Droid 1.0: Factory says its agent harness cuts one 312-tool task from 36 requests to 2 by calling tools programmatically (a vendor claim).
• Claude API limits: Anthropic quadrupled the request size ceiling to 128 MB on Messages and Token Counting, with higher per-image resolution, which helps image-heavy workloads.

Devtools & Infra

• Node.js 26 LTS: Vercel made Node.js 26 LTS generally available and the default for new projects.

Startups & Industry

• Reddit goes logged-in only: CEO Steve Huffman told shareholders Reddit is moving to a logged-in-only experience for better ad signals, and on the call said Google's AI Mode "does not drive traffic", with October search referrals down over 1 million a week. Yesterday's Alphabet sell-off had Google admitting AI Mode hurts classic search; today a major publisher confirms it from the receiving end.
• Anthropic's S-1 and Taiwan: A new detail from the filing: Anthropic says it depends heavily on TSMC chips made in Taiwan and warns of disruption from tensions with China, so its growth story carries a geopolitical single point of failure.
• DeepSeek's round: The Information says DeepSeek doubled its nearly $16B round because black-market Blackwell chip prices surged, with state-owned Chinese investors at 27%.
• AI debt: Microsoft plans to raise $25B in debt for its AI buildout, and Nvidia is reportedly negotiating to guarantee more than $5B of construction debt for Nscale's 460MW site. The buildout is increasingly financed by borrowing.
• Amazon on shopping agents: Andy Jassy told investors outside AI shopping agents are still not great, nor more convenient than going to a retailer.
• Chip demand: Intel says its budget Wildcat Lake laptop chips are sold out as RAM shortages push buyers to cheaper machines, while Sandisk reports revenue up 378% to $11B and an 8-for-1 split.
• Earnings misses: Pinterest guided Q4 below estimates and Twilio slowed its Q4 outlook; both fell after hours.

Hacker News

Binance dominated the front page: the hack tracker topped 1,000 points and 600 comments (discussion), covered in Security above. On the AI side, Claude Duet is praised as the first personal-assistant harness that feels useful, Proton's AI reckoning from 404 Media looks at internal fear and manifestos at the privacy company, and YC Labs ran agent swarms in Minecraft and Mindustry.

On the small-tools side, a Swift app auto-dismisses macOS's "find devices on local networks" prompt, a tiny world model runs games on a Raspberry Pi Pico 2 W, and an essay argues for the importance of communities around software.

Threads

• AI as attacker: ARTEX (state-linked espionage at scale) and Epoch's Firefox exploits (open-weight model, $1,000 each) show offense getting cheaper from both ends; the trojanized Claude Code extension and the Pentagon agent story show the agent tooling itself becoming the target.
• Binance and Polymarket: two exchange-scale security incidents in one day, one at a third-party provider and one in a hot wallet, with Binance covering losses and Polymarket reimbursing traders.
• Search traffic: Reddit's "AI Mode does not drive traffic" extends yesterday's Alphabet story, and its logged-in-only move is the publisher response.
• AI financing and supply: Microsoft's $25B debt, Nvidia's Nscale guarantee, DeepSeek's black-market chip pricing and Anthropic's TSMC risk all describe the same constraint: compute is expensive and geopolitically exposed.
• Credits and pricing: OpenAI matching Anthropic's API credits and Altman keeping token pricing both show labs competing on how developers pay.