Binance hit by a record $6.3B crypto hack

Binance was hit by a $6.3 billion hack, likely the largest crypto theft ever 🔒 Binance, the world's biggest crypto exchange, said on Oct 29 that about $6.3 billion in digital assets was affected, that about half has been frozen, and that it will "cover all losses." Binance co-CEO Richard Teng said early checks suggest it started with a compromised engineer's laptop, which gave the attacker access to a service container used to submit unauthorized withdrawals, and Binance later said it had found more points of breach. TRM Labs and Chainalysis both link the attack to North Korea's TraderTraitor hacking unit, and Binance has paused withdrawals while it checks its systems. Check Point's Jayant Dave said: "attackers no longer need the vault keys. They target the systems and people with authority to use them." Why it matters: your laptop holds the keys to your apps too, so keep production secrets off it where you can and turn on hardware security keys for the accounts that matter! (1,448 points on HN as of 4:40 AM PT Oct 30) BBC · TRM Labs · HN

Mistral Large 4 is now open weight, and it's built for coding! Mistral released the weights of Mistral Large 4, its 1-trillion-parameter multimodal model, and you can download them from Hugging Face and run them with vLLM or SGLang. Mistral says it's the best open-weight model for coding: it scores 48% on the Artificial Analysis Coding Agent Index, about tied with GLM-5.3 while using 47% fewer tokens per task, and 41% on Terminal-Bench 4.0. Those are Mistral's own numbers, so test it on your code before you switch. The license is a modified MIT license that lets you use, change and sell what you build, except that companies offering the model as a hosted service need a separate license. On Mistral's API it costs $0.70 per million input tokens and $2.10 per million output tokens. Why it matters: a strong coding model you can run on your own servers gives you more control over cost and data! (1,527 points on HN as of 4:40 AM PT Oct 30) Mistral · HN

Gemini 4 Carbon takes the #1 spot on a big AI leaderboard 🏆 Artificial Analysis, an independent benchmarking firm, says Google's Gemini 4 Carbon scores 59 on its new Intelligence Index v5, the top score. On its Coding Agent Index, Carbon with Antigravity CLI scores 59, just 3 points behind Claude Code with Claude Opus 5.5 at 62. It has a 1M-token context window and lists at $4 per million input tokens and $20 per million output tokens, with a 50% launch discount. The catch: Google is rolling it out to selected users only, so it isn't publicly available yet. Why it matters: Google is now close to the top for coding agents, which means more real choices for your coding tools soon! (336 points on HN as of 4:40 AM PT Oct 30) Artificial Analysis · HN

Martin Fowler says writers shouldn't let an LLM write their first draft. In an Oct 29 post, the well-known software author says that when prose starts to feel LLM-written, his desire to keep reading drops, and that even human editing struggles to remove that feel from an AI draft. He recommends authors write the prose themselves, says using an LLM for an outline may be OK, and asks anyone on his site who used AI to help write to say how they used it. He says he only uses AI as a search tool and checks every fact at the source. Why it matters: if you write docs, launch posts or READMEs with AI, readers may notice, so put your own words in! We're AI-drafted ourselves, which is why we say so in every issue. (499 points on HN as of 4:40 AM PT Oct 30) Martin Fowler · HN

A dev group says Claude Code burns tokens on 3-word status lines, and Anthropic responds 🍿 In a public gist, Flamerix AI says that about every 30 seconds Claude Code sends each background agent's full transcript to its model just to write a 3-to-5-word status line in the agents panel. In its own telemetry, it says this took 16.4% of its Claude Code spend, there's no setting to turn it off, and it published a script that blocks the timer, while warning that an update can undo it and that changing the client may conflict with the service's terms. A member of the Claude Code team replied on HN that only actively working background agents get these calls, and that across users who send telemetry it was about 2% of cost at API list prices last week, or about 6% in the terminal. The team member agreed that's a lot for a status line and said Anthropic just rolled out a fix for the most expensive case in version 2.1.311 and later, and is building a much cheaper mode. Why it matters: if you run lots of background agents, update Claude Code and keep an eye on your usage! (567 points on HN as of 4:40 AM PT Oct 30) Flamerix AI gist · Anthropic's reply on HN · HN