The day's tech, sifted: Nov 01, 2026

What matters today: Anthropic's IPO filing is public and the commentary has started: its S-1 in charts sits next to warnings of cracks in the AI spending boom. The Binance hack has an author: the FBI says North Korea's TraderTraitor group did it, and Binance says it will resume withdrawals in phases.

Security & Privacy

• Binance, the delta: Yesterday Binance said it had covered 90% of the loss; it now says 95% of the ~$6.3B in stolen tokens is covered and withdrawals restart in phases from Sunday. The FBI attributes the theft to TraderTraitor, the North Korean group it also blamed for the ~$1.5B Bybit hack, and Elliptic says the laundering is running on autopilot, so recovery of the rest looks unlikely.
• ChatGPT account takeover: Parameter Labs chained a sandboxed-iframe escape in ChatGPT's HTML previews (a virtual-keyboard message handler that inserted raw HTML, then a YouTube endpoint that reflected script past the CSP). A crafted share link could read a signed-in visitor's private data. OpenAI fixed it and paid a $5,000 bounty, which Simon Willison calls cheap for the severity.
• Backdooring Bash: Hacktron describes how the GNU Savannah hosting setup could have been used to backdoor Bash, a reminder that a project everyone depends on can sit behind thinly guarded infrastructure.

AI / LLMs

• Chatbots and midterm facts: The Washington Post found chatbots lean on three obscure AI-built voter sites for US midterm answers; Copilot cited the error-prone Ballot Brief in 75% of tests and ChatGPT in 22%. Voters asking an assistant may get answers from a single unvetted source.
• DeepSeek Super-Agent: DeepSeek open-sourced Super-Agent as a research prototype, a proactive coding agent that can watch group chats for coding issues and suggest fixes, aimed at Claude Code's popularity.
• Agent playtesting: Aura open-sourced Autopilot, an in-editor agent that playtests Unreal, Unity and Godot games from plain-English prompts; Aura claims it beats Claude Code on accuracy and speed in a seven-task benchmark of its own.
• One photo to 3D: Sperid Labs released Idol, an Apache 2.0 model that turns one image into up to 1M Gaussian splats (a point-based 3D format) in seconds on consumer GPUs.

Devtools & Infra

• Datasette Agent: datasette-llm 0.1a9 now reads the default model from the LLM tool, so Datasette Agent runs without extra model configuration.

Startups & Industry

• Anthropic's IPO, read two ways: Beyond the S-1 charts, SemiAnalysis argues the IPO is the moment datacenter developers have been waiting for, while the BBC asks whether little-known firms feed fears of an AI bubble. The money story from yesterday continues: filing detail now meets the doubt.
• Insurance AI deal: Guidewire agreed to buy Hyperexponential for $1.7B in cash (AI underwriting and pricing software for insurers), plus up to $400M more per Reuters.
• Voice AI consolidates: Deepgram bought Daily, maker of the open-source Pipecat voice-agent framework, for an undisclosed sum.
• Funding: Lumaril raised $66M at a $400M valuation for AI that tunes factory lines through existing control systems; Shanghai's NeuroXess raised about $105M for brain-computer interfaces.

Elsewhere

• Apple after apps: Fast Company reads HomeView as a preview of a Siri-first, post-app Apple.
• Robotics: Robonomics looks at robotics after Astra.

Hacker News

The sharpest thread was a Linux flaw: a TCP use-after-free rated CVSS 9.8 (432 points, 196 comments), a score meaning remote, easy exploitation. Anthropic drew two critical stories: users say it blocks anything touching training larger models (282 points, 158 comments), and it is blocking Claude from reverse-engineering video games (125 points). On capability claims, a post says GPT-5.6 Sol can beat Factorio on peaceful mode (178 comments), and a repo claims five Qwen3.8-27B models beat Fable 5 on LiveCodeBench Hard; treat both as unverified single-author results.

For systems readers, Jepsen tested ParticleDB 1.1.4 (234 points), and a memory-latency essay argued latency, not bandwidth, is the limit. The crowd favorite was HackerTube (974 points, 299 comments), a YouTube-style browser over 6,454 videos linked in HN comments since 2007.

Threads

• Anthropic under the microscope: the S-1 charts, the datacenter-demand argument and the Hacker News complaints about training and game-cloning limits all concern what Anthropic will and will not let customers do as it goes public.
• AI money, doubt and deals: cracks in the spending boom and the BBC bubble piece sit against Guidewire, Deepgram and Lumaril still closing AI deals.
• Coding agents multiply: DeepSeek Super-Agent and Aura Autopilot both pitch themselves against Claude Code.
• Trusted surfaces fail: the ChatGPT preview exploit, the Bash hosting risk and chatbots citing weak voter sites show defaults being trusted past what they earn.