AI news · Sunday, November 1, 2026

Hackers hijacked a fake Claude coding extension for 11 hours

A tampered version of Anthropic's Claude Code extension sat on Cursor's marketplace for about 11 hours, secretly trying to install a backdoor on developers' computers using remote-access tools. Cursor removed it and says flaws in the hackers' code likely stopped most attacks from working. Anthropic says only Cursor users were affected, not its own systems.

The catch: As AI coding tools become standard in software shops, a single poisoned extension can reach thousands of developer machines before anyone notices.

via Fast Company

Researchers broke into ChatGPT accounts with just a shared link

Security researchers found a flaw letting an attacker take over any ChatGPT account just by sending a victim a link to a shared conversation. Opening it ran hidden code that could read private chats and call ChatGPT's own systems as the victim, no download or click needed beyond opening the link. OpenAI fixed the bug on October 26 and rated it its most serious severity level.

The catch: The bug came from two harmless features interacting in an unexpected way, a reminder that AI products are only as safe as their weakest combination.

via Parameter.ai (via Hacker News)

Mouse brain surgery, done by an AI agent with a robotic arm

A biotech startup let an AI agent choose a gene therapy, pick injection coordinates, and control a robotic needle to inject a live mouse's brain, with a vet surgeon doing the actual cut. Anthropic's Claude Opus model refused mid-task, saying something told it not to proceed; the team switched to Claude Sonnet, which completed the injection with human oversight and an emergency stop button.

The twist: The more cautious model refusing is as notable as the surgery itself: safety behavior and capability don't always point the same direction.

via Wired

OpenAI reportedly cuts funding target from $30B to $20B

OpenAI has lowered its funding goal for its newest round from a previously reported $30 billion to $20 billion, with the first close expected even lower, the Financial Times reports. The round is being led by UAE investors; the Wall Street Journal says OpenAI has secured $11 billion so far and gotten a "tepid" response from some big investors, amid broader Wall Street wariness about AI infrastructure spending.

Follow the money: When the industry's most-hyped company has trouble filling a funding round, it's a sign investor patience with AI's cash burn is thinning.

via TechCrunch

Anthropic may get a 'junk' credit rating after its IPO

Blackstone, an anchor investor in Anthropic's chip financing, now expects the AI company to receive a high "junk" credit rating after going public, the Financial Times reports, reversing an earlier prediction of an investment-grade rating by 2027. A junk rating would make it costlier for Anthropic to borrow money to keep funding its heavy spending on computing power.

Follow the money: A worse credit rating raises Anthropic's borrowing costs right when it needs to keep raising huge sums to pay for AI computing power.

via TechCrunch

Fujifilm is quintupling old-school tape production for AI data

Fujifilm will spend about $63 million to boost production of magnetic data tape fivefold by 2028, saying generative AI is driving "exponential growth" in how much data companies need to store. Despite seeming outdated, tape is cheap, can be physically locked away, and is increasingly used by cloud companies as a backup defense against ransomware attacks.

Zoom out: AI's data boom is reviving 1950s-era storage technology because nothing else beats tape's combination of cost and offline security.

via Ars Technica

Shopping AI agents are live, but retailers say use is tiny

Meta's shopping agent Muse has over 10 million activated US and Canada users, and Zuckerberg says 25% use it to shop. But in earnings calls, DoorDash called agent orders "hundreds," Hilton called bookings "insignificant," and eBay said activity was "very, very small." Most retailers say they expect to benefit rather than be disrupted, for now.

The catch: The gap between Meta's splashy usage numbers and retailers' shrugging suggests AI shopping agents are still mostly a novelty, not a habit.

via Wired

Malware now uses AI to automatically profile its victims

Security researchers at Cisco Talos found infostealer malware, nicknamed Volta-Autodoxer, that feeds stolen data like passwords, Discord logins, and food-delivery receipts into AI chatbots to build a detailed profile of each victim automatically. It defaults to Anthropic's Claude, with fallbacks to other AI models if one refuses the request, using a prompt written to avoid triggering safety filters.

What's next: As AI safety filters get better at blocking obviously malicious requests, attackers are just building in backup models that will say yes.

via Wired

Otter.ai wants to tell you what to say in your next meeting

Otter.ai, the transcription app used by over 40 million people, is adding features that go beyond summarizing meetings to suggesting what participants should say next and auto-preparing briefing notes before meetings start. Paid plans run from $99.96 to $239.88 per user a year. The company says the real goal is fixing meetings that are inefficient because people show up unprepared.

For you: If your workplace already uses Otter for notes, this is the kind of quiet feature creep that turns a transcription tool into a daily coach.

via Fast Company

How to turn off Apple's pushy AI writing button

Many iPhone and Mac users have been annoyed by Apple's new "Write with Siri" button, which appears above the keyboard by default in iOS 27 and can't be turned off from the normal Siri settings. The fix: go to Settings, Screen Time, Content & Privacy Restrictions, turn that on, then tap Siri, Writing Assistance, and choose "Don't Allow" to remove the button entirely.

For you: A genuinely useful fix for anyone who finds AI writing suggestions more distracting than helpful, and a sign these tools are being forced on by default.

via Fast Company

What it means when an AI 'hallucinates'

When an AI chatbot "hallucinates," it confidently states something false as fact. AI models don't look things up; they predict likely-sounding words based on training patterns, which can produce fluent but wrong answers. For example, asked for academic sources, a chatbot may invent realistic book titles, authors, and page numbers that don't exist. The confident tone makes errors easy to miss, so always verify AI-given facts independently before relying on them.

For you: Knowing this turns you from someone who trusts AI answers by default into someone who knows exactly when to double-check them.

Original explainer

Judge lets pollution lawsuit against xAI's Memphis site proceed

A federal judge rejected xAI's attempt to dismiss an NAACP lawsuit alleging the company ran at least 35 gas turbines at its Colossus data center near Memphis without required air pollution permits, a possible Clean Air Act violation. xAI argued the turbines were exempt as "mobile sources" and that plaintiffs lacked standing; the judge disagreed on both points, so the case will now proceed.

What's next: As data center backlash grows nationally, this case will test whether AI companies can be held to the same pollution rules as any other industrial operator.

via TechCrunch