A backdoored Claude extension, Anthropic's agent-era review, and $828B of capex

The Pulse — Gergely Orosz's weekly reports three structural moves at once: Anthropic has removed the algorithmic coding interview from its engineering loop as signal-free, Amazon has written AI usage into formal expectations for engineers and their managers after correlating it with productivity, and Microsoft has offered 60,000 employees exit packages by year-end. Hiring and evaluation processes built for hand-written code are being reworked first at the companies setting the pace — worth knowing before your own interview loops and performance rubrics come up for review. The measured number in the issue: incident.io cut its AI bill from $4,500 per developer per month in August to $1,600 now. The Pulse (Gergely Orosz)

Claude Code review — With Claude writing roughly 90% of Claude Code, Anthropic has documented how review holds together: a classifier assigns each change a risk tier, low-risk changes merge with no human approval, and an agent's sign-off never substitutes for a person's. Daily merged changes grew from about 2,500 in May to over 30,000 in September. It is the most detailed public template for agent-era review, including the failure mode that forced a fix: one CLI change spent a night shuttling between agent author and agent reviewer for over fifty cycles before a human killed the run, and running more rounds by default made output worse, so the team now caps them. InfoQ

“Nobody on the team keeps a complete mental model of Claude Code in their head, and we don’t think anyone needs to.”

Atlassian agent governance — Atlassian's argument is that unmanaged agents look like sanctioned users holding sanctioned credentials, which makes sprawl an identity and accountability risk rather than ordinary shadow IT. Its proposed controls are permission inheritance, per-agent scopes, and an admin-facing inventory of what is running and who owns it. If your organisation is past its first hundred agents, someone needs that inventory before an incident or an auditor forces the question. DocuSign, the worked example, cut ticket drafting from about 20 minutes to 5 and scaled standard agents across more than 100 Confluence spaces; Gartner expects a Fortune 500 to run over 150,000 agents by 2028. Atlassian (Inside Atlassian)

DoorDash — On its third-quarter call, DoorDash said third-party agents like Muse have produced only several hundred orders in total, against roughly 3 million deliveries a day. CEO Tony Xu says he will not fight the trend, but will not be delusional about its current size either. A useful calibration when the board asks when agents will transact with your product: the connector plumbing shipped only weeks ago, and DoorDash's own in-app agent — whose baskets run nearly 50% higher than traditional orders — is where the volume actually sits. Amazon blocks personal agents from ordering; Walmart has welcomed them. TechCrunch

Gemini 4 Carbon — Google's strongest model, aimed at security work and multi-day engineering tasks, arrived two weeks after Argon — but only for vetted cyber-defence teams at first, widening through its Fairwind Program before general release. Plan around what you can actually buy, not what tops the rankings this week; the durable signal is that cyber-capable frontier models now get gated rollouts, so your security team should be applying to those early-access programs. The introductory API price is $2 per million input and $10 per million output tokens, doubling to $4/$20 after the launch window; cached input is 95% off. Google blog

GitHub Copilot — At Universe, GitHub pitched itself as an agentic software delivery system: a Copilot app that runs work from issue to merge, automatic model routing under Fusion, and parallel agent projects across repositories. The commercial change matters more than the demos — Enterprise customers now get every Copilot feature without buying a separate subscription, which moves the decision from seat counts to usage metering. Existing Copilot Business and Enterprise customers can stay on their current plans or migrate; customers pay only for the AI credits they burn. GitHub Blog

Claude Ultra — Anthropic renamed its Max plans — the $100 tier is now Max, the $200 tier is Ultra — and removed the five-hour rolling usage limit for verified Ultra accounts, starting with most existing subscribers. If your heaviest agent users were shaping their work around the reset window, the top tier now sustains long-running agent jobs without the mid-session cliff. No subscriber action is needed, and prices are unchanged. explainx.ai

Hyperscaler capex — Amazon, Alphabet, Meta, Microsoft and SpaceX spent about $227 billion on capital expenditure in the third quarter, roughly 20% more than the quarter before, and guidance puts the full year between $803 billion and $828 billion. Amazon was the largest spender at $75.9 billion, mostly AWS and AI. This is the bill behind the compute your agent vendors rent, and the number to quote when the board asks whether AI prices can keep falling. All five have signalled spending rises again in 2027. Business Insider

AI spending wobble — Alphabet posted its second consecutive quarter of negative free cash flow, minus $9.1 billion, and its shares fell 12% in a day, erasing $500 billion of market value; Amazon's trailing free cash flow is negative $42.4 billion. The big spenders plus Oracle raised $326 billion of debt this year, triple 2025. This matters to you as counterparty risk: the FT reports Blackstone warns Anthropic may not earn an investment-grade rating after its IPO, tying the vendor closer to hyperscaler financing. Microsoft was the exception, at $18.6 billion of positive free cash flow, though it blamed data-center delays rather than restraint. TechCrunch

“It’s spending billions on the AI technology that could undermine the search business that pays it.”

OpenAI funding round — OpenAI has lowered the target of its current round from $30 billion to $20 billion after securing about $11 billion, and its valuation ambition from $1.4 trillion to $1.25 trillion. The Wall Street Journal reports some large investors turned tepid; a UAE consortium is expected to lead, and the first close may land under target. The read-through for you: the two biggest agent vendors now compete for the same institutional money, and the FT reports OpenAI expects to burn $278 billion of cash through 2030 — pricing power at the model layer may be softer than list prices suggest. OpenAI plans to list as early as next year; Anthropic is expected to price its IPO in mid-November. Trending Topics

Epoch patch-gap research — Epoch AI showed the open-weight GLM 5.3 turning the gap between a fix landing upstream and reaching users into working exploits against Firefox versions released between September 17 and October 26 — each compromise costing about $1,000 of API spend, reading local files from a malicious page with no safeguards triggered. Mozilla patched the undisclosed bug on October 27. Your patch latency on internet-facing software is now a window a commodity model can walk through unattended; a monthly patch cadence does not cover it. The same effort failed against Chrome despite roughly $100,000 spent on attempts. Epoch AI

Anthropic S-1 — Anthropic filed its S-1 to list on Nasdaq as ANT, in a debut reportedly worth up to $100 billion; the filing shows an $80.7 billion first-half loss on $16.3 billion of revenue — the yardstick for every AI vendor negotiation next year. Pricing is expected mid-November, reportedly at $1.5 trillion or more. Anthropic

GPT-6.1 Terra — OpenAI's new mid-tier model handles routine computer-use work — reading screens and typing into apps — at $1/$5 per million tokens, half its predecessor's price; a candidate for high-volume agent tasks once independently measured. Already generally available in GitHub Copilot. The Next Web

ARTEX — CrowdStrike says China-linked hackers used an AI tool built on a DeepSeek model to hit over 200 organisations in 11 days, compromising at least nine — the clearest sign yet that AI is compressing intrusion timelines. Governments and businesses worldwide were among the targets. NBC News via Techmeme

Claude exploit research — Anthropic says an internal research model built working exploit chains across Chrome, Safari, Android, Windows and Linux, leading to 62 patched Chrome CVEs — offensive capability once held by a handful of specialist teams. All findings went through coordinated disclosure; patches are shipping. Anthropic

Genie 4 — DeepMind's Genie 4 generates real-time interactive worlds with native audio and dialogue, out to trusted testers as a research preview — not a product decision this quarter, but a pointer to where agent training environments are heading. All output carries SynthID watermarking. Google DeepMind