The day's tech, sifted: Oct 31, 2026

What matters today: The Claude Code extension scare from yesterday now has an Anthropic statement: a malicious actor used access to Cursor's marketplace to publish a tampered copy, and Anthropic's own systems were not compromised. Money got pickier on the same day: OpenAI trimmed its round to about $20B at a $1.25T valuation, and Intel and Reddit both had their worst days in years.

Security & Privacy

• Claude Code extension, the update: Yesterday's backdoor report is now pinned on Cursor's marketplace. Ars says version 2.1.313 was live for nearly 12 hours and reached at most 5,000 users, and researchers say it went out through Anthropic's official account there; Anthropic says it did not build or publish the copy, and neither company has explained how the account was reached. Anyone who ran 2.1.313 should treat that machine's credentials as exposed.
• Wiz on a "Black October": Wiz frames this month's incidents at AI developer platforms as a pattern, so the extension compromise reads as one of several attacks on the tools developers trust by default.
• Shared AI opt-out: OpenAI, Google, Microsoft and Anthropic say they will adopt a single robots.txt signal for AI training preferences, proposed at the IETF (the internet standards body) by an Anthropic standards lead. Publishers set one rule instead of one per crawler; Google's note confirms support.
• VPNs in Russia: EFF says VPN apps are again being pulled from Apple's App Store in Russia, cutting off a main route around state censorship.
• Scribd notice: A user posted an emailed notice that Scribd had a security incident affecting its brands, including Slideshare and Everand; no scope is public yet.

AI / LLMs

• Claude plan change: Anthropic renamed Max 5x to "Max" ($100/month) and Max 20x to "Ultra" ($200/month), and verified Ultra accounts lose the 5-hour session limit; weekly limits stay. Heavy users get uninterrupted sessions.
• Teen weapons limits: OpenAI says ChatGPT will give teens less weapons-related information, even for questions that look harmless.
• Agents and food delivery: DoorDash's CEO says outside AI agents have produced only several hundred orders so far, a reality check on agentic shopping.
• Hallucination benchmark: Artificial Analysis cut AA-Omniscience to 3,000 questions and says it now exposes models that guess instead of abstaining.
• Altman on RAM: Sam Altman says it is better for RAM to sit in data centers than in personal PCs, a stance that lands badly amid a memory shortage.
• Dartmouth provost: Provost Santiago Schnell resigned after a committee found undisclosed AI use in a paper and an op-ed on AI cheating.

Devtools & Infra

• Pydantic 3: Pydantic V3 is in development, V1 reaches end of life and a paid enterprise support tier arrives; teams still on V1 need a migration plan.
• Nesformer: Drew Breunig baked a working NES CPU emulator into the weights of Qwen3.5-4B, so Super Mario Bros runs, very slowly, inside a language model.

Startups & Industry

• OpenAI's raise shrinks: The target fell from $30B at $1.4T to about $20B at $1.25T, and the WSJ says only about $11B is committed, less than half the original goal. It lands just before Anthropic's IPO, which TechCrunch reports is expected to draw a high "junk" credit rating.
• Debt gets pricier: S&P moved Meta's outlook to negative over capex and negative free cash flow, and the WSJ says CoreWeave, Nebius and Nscale are effectively shut out of the bond market. Big Tech has now committed over $3 trillion to AI build-out, and the cheap money behind it is thinning.
• Google's AI Mode: Google's business chief said AI Mode's growth came with lower monetization, tying the sell-off to the summary-versus-source tension publishers keep raising.
• Intel and Reddit: Intel fell 28.6%, its worst day since the 1970s, on weak guidance, and Reddit dropped 29% after a weak Q4 forecast.
• Baseten: Baseten is raising about $2B at $26B and $30B valuations, up from $13B in June, evidence that inference infrastructure still draws money.
• Apple Intelligence in China: Apple will bring earlier Apple Intelligence features to China with iOS 27.3, beta November 4 and release in December, without the new Siri.

Elsewhere

• Copyright and AI: EFF filed an amicus brief in In re OpenAI arguing that rewriting copyright law for generative AI would hurt researchers, artists and the public, while OpenAI publishes supportive briefs on fair use.
• Hubble: NASA is again asking for ideas to boost Hubble's orbit; left alone it re-enters in the early-to-mid 2030s.

Hacker News

Infrastructure trust was the front-page worry: two providers posted incidents, a Bunny CDN security incident (212 points) and a Linode security incident (176 points, 97 comments), each with little detail in the notice itself. The Scribd breach notice above also made the page.

On work and AI, Cisco's president says he gave a 27-year-old engineer unlimited AI tokens to rewrite legacy code, which drew a long argument about entry-level jobs (178 comments). A French National Assembly vote on a per-gigabyte internet tax split readers (123 comments). The warmest thread was the Boston Globe's Patrick Stewart tribute (267 points), with a TNG guessing game alongside, and the cold one was Stone Age finds emerging from melting ice.

Threads

• Developer-tool supply chain: the Claude Code extension update and Wiz's "Black October" show attackers going after marketplaces developers auto-update from; the Bunny and Linode incidents on Hacker News add infrastructure vendors to the list.
• AI money tightens: OpenAI's smaller raise, Meta's outlook, the neocloud bond freeze and Anthropic's expected junk rating are one story about the cost of capital; Baseten is the exception.
• Who gets paid for content: the shared AI opt-out, Google's AI Mode revenue admission and the EFF copyright brief all circle publishers versus AI summaries.
• Agents meet reality: DoorDash's few hundred agent orders and the Dartmouth provost's undisclosed AI use sit against the Cisco unlimited-tokens bet on AI doing the work.