AI News Today — October 31, 2026: A Backdoored Claude Code Extension, 'Several Hundred' Agent Orders, and the AI Spending Selloff

The big picture

Halloween delivered the scary version of the AI story: trust and money. A tampered copy of Anthropic's Claude Code extension sat in Cursor's marketplace for nearly 12 hours planting an SSH backdoor on developers' machines. DoorDash gave agentic commerce its first hard number — several hundred orders from outside AI agents, against 993 million orders in one quarter. And investors spent the week doing the math on the buildout itself: the five biggest U.S. tech firms' capital spending has now outrun the cash their businesses generate, Nvidia is reportedly weighing cosigning a customer's construction debt to keep one project financeable, and Intel's crash alone knocked $12 billion off the U.S. government's stake. The technology keeps advancing; October 30 was about whether anyone can trust it — and pay for it.

What happened today

A backdoored Claude Code extension reached up to 5,000 Cursor users

What happened: Cursor disclosed that malicious versions of Anthropic's Claude Code extension — 2.1.312 and 2.1.313 — were distributed through its VS Code extension marketplace on October 29 between about 2:27 a.m. and 1:35 p.m. Pacific Time, affecting at most 5,000 users on macOS and Windows. The packages were not developed by Anthropic. On install, the code collected host and user details, tried to add an attacker-controlled SSH key labeled svc-telemetry to ~/.ssh/authorized_keys, and installed persistence — a LaunchAgent named com.apple.sysmond on macOS, a scheduled task on Windows — retrying a reverse tunnel roughly every 60 seconds. Cursor says the SSH-enablement step would normally fail without administrator privileges, removed the versions, shut down and rebuilt its marketplace, rotated its keys, and had the attacker server taken offline. There is no evidence the official Microsoft VS Code Marketplace or Open VSX copies were affected, and Anthropic has published version 2.1.314.

Who it matters to: Any developer who uses Cursor — and, by extension, anyone whose code, cloud credentials or repositories sit on a machine that does. Developer workstations are the highest-value target in the building, which is exactly why the attackers picked an AI coding extension as the vehicle.

What you can do: If you installed or updated Claude Code from the Cursor marketplace on October 29, check ~/.ssh/authorized_keys for a key labeled svc-telemetry and remove it, look for the persistence items Cursor names in its notice, and rotate credentials reachable from that machine. Uninstalling the extension alone is not enough — the payload drops standalone background scripts. Then update to 2.1.314 from an official marketplace.

Learn more: Ars Technica on the backdoored extension and Cursor's official security notice

DoorDash has received 'several hundred' orders from AI agents — total

What happened: On DoorDash's third-quarter earnings call, CEO Tony Xu told analysts the company has seen "maybe several hundred orders in total from some of these third-party agents" — naming Meta's Muse, Instinct, Fo "and others." The comparison number is DoorDash's own: 993 million orders in Q3, with revenue up 34% to $4.6 billion. Xu's framing was measured — "you never want to fight gravity. But at the same time, you don't want to be delusional" — and he noted most ordering still happens in DoorDash's own app, where the company runs its own Ask DoorDash, text and CLI agents. The figure counts third-party agents only, and those agents are weeks old; it is a starting line, not a verdict. But it is the first scale measurement anyone has put on a year of agentic-commerce claims.

Who it matters to: Anyone building, selling or investing on the thesis that assistants replace apps as the front door to commerce. The infrastructure for agent ordering exists — DoorDash even ships an MCP connector for it — and, so far, almost nobody is using it.

What you can do: If you're a local business, keep your effort on the channels customers actually order through today, and treat agent-readiness — clean menus, structured data, working connectors — as cheap insurance rather than a growth bet. If you're evaluating agent products, ask vendors for transaction counts, not demos.

Learn more: TechCrunch on DoorDash's agent orders

The AI spending boom showed cracks — in the cash-flow numbers

What happened: A Reuters analysis of the five biggest U.S. tech firms — Microsoft, Alphabet, Amazon, Meta and Oracle — found operating cash flow rose 29% to a record $542.9 billion in the first nine months of 2026, but capital spending, driven by AI, doubled to $556.9 billion, exceeding the cash generated for the first time in data going back to 2011. The combined free-cash-flow result swung from a $143.5 billion surplus a year earlier to a $14 billion deficit; buybacks fell 83% to $13.6 billion, and the share of cash returned to shareholders dropped to 9.3% from 27.4% — the smallest in at least 15 years. Markets reacted hard: Alphabet fell 12.6% in one session, wiping out roughly $500 billion in market value, and Intel plunged after its own cash-burning results, knocking $12 billion from the government's stake. In the words of one analyst quoted by Reuters, the market "isn't necessarily questioning whether the AI buildout can be financed — it's increasingly questioning whether shareholders will earn an adequate return."

Who it matters to: Everyone downstream of Big Tech pricing. When the builders' cash flow goes negative, the pressure shows up as higher cloud and device prices, thinner free tiers and harder selling of AI subscriptions — trends this briefing has tracked all month.

What you can do: If your business runs on AI subscriptions or cloud credits priced in the subsidy era, model a paid, full-price version of your stack now. The era of investors cheerfully funding below-cost AI is what is actually ending.

Learn more: TechCrunch on the cracks in the AI spending boom

Nvidia is reportedly weighing cosigning a customer's data center debt

What happened: TechCrunch reports, following The Information and corroborated by Bloomberg and Reuters, that Nvidia is in talks to guarantee billions of dollars of debt for Nscale's planned data center campus in West Virginia. Anthropic has signed a $44.6 billion contract for the project's first building alone, a 460-megawatt facility expected to run Nvidia's next-generation Vera Rubin chips, according to Nscale's IPO filing. The proposed structure is a residual value guarantee: if Nscale defaulted on construction payments, Nvidia would step in, temporarily take over the facility and help find a buyer and new tenant. Nvidia's previous support for cloud providers — roughly $170 billion in investments and credit support, per The Information, including $1 billion of convertible financing committed to Nscale in September — mostly covered leases and chip purchases, not construction. No agreement has been signed and terms could change; treat the deal itself as unconfirmed.

Who it matters to: Anyone trying to read the buildout's financial health. When the chip supplier also becomes the landlord's cosigner, the same balance sheet is carrying the project at three points at once — which is precisely the circularity investors started punishing this week.

What you can do: For now, nothing transactional — but file this next to the cash-flow story above. Together they are the clearest picture yet of how the AI buildout is actually funded, and where the risk concentrates if demand disappoints.

Learn more: TechCrunch on the Nvidia–Nscale talks

Muse AI: Meta's agent is quietly gaining a signed ID — reported, not announced

What happened: Our daily priority check found no new Muse launch or feature announcement from Meta. The one fresh development is a detection, not a release: DataDome VP of Threat Research Jérôme Segura reported on October 30 that Meta has been attaching cryptographic signatures — the Web Bot Auth protocol — to a growing number of web requests from its Muse personal agent and its Meta-WebIndexer crawler since around October 12, in what appears to be a quiet, small-volume test. A signed request lets a website verify that an automated visitor really is Meta's agent rather than an impersonator. Meta has not announced the change, and its public crawler documentation does not yet mention Web Bot Auth. The context matters: in September, Amazon blocked Muse, saying the agent did not identify itself — verifiable identity is becoming the admission ticket for agents on the open web.

Who it matters to: Website owners and publishers first — signed agent traffic is how you will soon tell welcome agents from scrapers wearing their name. It also matters to Muse users: better identification is what gets agents un-blocked from the stores and sites they are supposed to act on.

What you can do: If you run a website, watch your bot logs for signed Meta agent requests and decide your policy for identified agents now, rather than after one shows up in your checkout. If you use Muse, there is nothing to switch on — this is infrastructure arriving underneath the product.

Learn more: PPC Land on Muse's signed ID test

Gemini Spark: no new announcement — Task mode stands, 17 days to the Gems migration

What happened: Our daily priority check found no new Gemini Spark announcement. Spark remains folded into the Gemini app as Task mode — the Chat | Task switch above the prompt box, a Scheduled page grouping Active, Paused, Completed and Suggested tasks, and Task settings in the account menu — for Google AI Pro and Ultra subscribers on personal accounts. The next dated change is now 17 days out: on November 17, Google begins automatically converting personal-account Gems into Skills, which Gemini can also apply on its own when it judges one fits your prompt. Work and school accounts follow in 2027.

Who it matters to: Anyone who built custom Gems or relies on Spark tasks to run errands in the background. The rebrand is finished; the migration of your saved setups is the event that can actually break something you depend on.

What you can do: This weekend, inventory your Gems: which ones you use, what each is supposed to do, and any instructions you'd hate to lose. After November 17, open each converted Skill once and test it before trusting it with a real task.

Learn more: FindSkill's explainer on Spark becoming Task mode and Android Police's field guide to the Gemini renames

Also worth knowing

• Anthropic is expected to get a high 'junk' credit rating after its IPO, the Financial Times reports — a formal label for the cash-burning economics underneath the week's selloff. TechCrunch
• Amazon's CEO said the same thing as DoorDash's. On Amazon's call the same day, Andy Jassy said few customers are shopping through outside agents yet — two of the biggest consumer ordering businesses, one message. FourWeekMBA's summary
• Singapore got its first agent-first store. Retailer Tanatasa announced a store built for personal AI agents — Muse, OpenAI's Dots, Instinct and Claude Duet among them — to search, build carts and prepare purchases, days after Muse arrived in Singapore. Kaka Computer's report
• Zest's new agent books restaurants over text. It finds restaurants and makes reservations by message — the single-task, repeatable end of the agent market, launching the same week DoorDash measured how small that market still is. TechCrunch
• Ollie, a privacy-focused AI assistant, is shutting down. One of the small independents that competed on privacy rather than scale is gone — the consumer assistant market keeps consolidating around the platforms. TechCrunch
• Google's first data center satellite is online — and writing poetry. An early orbital-compute experiment, notable as a signal of how far the search for data center capacity and power is willing to go. TechCrunch
• Can data centers be mass-produced like cars? A startup is aiming to find out, pitching factory-built facilities that come online in as little as 48 hours — aimed squarely at the construction bottleneck behind stories like Nscale's. TechCrunch
• A Theranos whistleblower's warning for the AI boom. Tyler Shultz, who helped expose Theranos, warned that the AI boom is "a great opportunity to go deceive people" — worth hearing on a day when a backdoor shipped inside a trusted tool. TechCrunch

One practical takeaway

Spend ten minutes on agent hygiene today. If you code in Cursor, run the svc-telemetry check in the first story above — it is the rare security action that is both urgent and quick. And wherever an AI agent acts for you, apply the DoorDash lesson in reverse: agents are real but tiny and new, so give them identified, reversible, low-value tasks — and verify, the way websites soon will with signed IDs, that what acted was actually your agent.

Sources

Cursor's official security notice and Ars Technica (Claude Code extension backdoor); TechCrunch and DoorDash's Q3 release and earnings call (agent orders); Reuters analysis of LSEG data and company reports, syndicated, and TechCrunch (AI spending, Intel); The Information via TechCrunch, Reuters and Bloomberg (Nvidia–Nscale talks — reported, not signed); DataDome via PPC Land (Muse signed ID test — reported, not announced by Meta); FindSkill and Android Police (Gemini Spark Task mode and the November 17 Gems-to-Skills migration); Financial Times via TechCrunch (Anthropic rating); TechCrunch AI feed (Zest, Ollie, Google satellite, mass-produced data centers, Tyler Shultz). Summaries are original; follow the links for full reporting.