A video description from AI LABS says Anthropic’s Claude Code team shared a workflow for checking its AI workers’ output.͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏
THE FRONTIER BRIEF
2026-10-31 · Signal, Not Noise
■ The Big Story
A new way to check AI-written code meets a security scare
A video description from AI LABS says Anthropic’s Claude Code team shared a workflow for checking its AI workers’ output. It describes a write-evals skill that turns each feature into checks and a design mockup, then blocks building until someone reviews them. The description reports more than twice as many changes and a drop of over 40% in the share that had to be undone. The video description links to an Anthropic article, but its text is not included here.
The timing is awkward: Ars Technica reports that an official Anthropic coding extension on Cursor carried a backdoor for nearly 12 hours. The compromised version, 2.1.313, affected at most 5,000 users. Anthropic pushed version 2.1.314 several hours later, according to researcher Kevin Beaumont. A review gate can catch flawed work; it cannot protect a team from a compromised tool used to do that work.
Why It Matters for Business Leaders
If your team uses AI to write software, treat reviews of its output and control over installed extensions as two separate safeguards.
OpenAI offers websites a shared way to state training preferences[OpenAI]
OpenAI says it will honor a shared robots.txt signal through which website operators can communicate AI training preferences to participating providers. It aims to implement the signal in the coming months, alongside its existing controls.
In Plain English: A website could post one preference notice for multiple participating AI companies to read, rather than addressing each company separately.
Why it matters: Website owners should keep their current settings in place while preparing for the additional option.
Zest puts restaurant recommendations in text messages[TechCrunch]
TechCrunch reports that Zest introduced a personal AI-powered restaurant concierge that lives in text messages. Its founder says replies draw on a user’s history and their friends’ histories.
In Plain English: The restaurant suggestion arrives in a familiar text conversation, informed by places the customer and their friends know.
Why it matters: Restaurants have another reason to consider how they appear in customers’ recommendation and booking journeys.
DoorDash puts AI-driven ordering in perspective[Ars Technica]
DoorDash CEO Tony Xu said outside AI workers had generated several hundred orders on DoorDash. Ars Technica reports that DoorDash handled just shy of 1 billion orders in the third quarter of 2026.
In Plain English: Software placing an order for a customer exists, but it accounts for a tiny fraction of orders at this major delivery service.
Why it matters: Merchants should judge AI-ordering proposals against actual customer demand, not the size of the promised market.
■ The Street Says
The discussion spans software-tool security, AI-company financing and whether paid AI products earn their place in everyday work.
↳ Re: backdoors in official Claude Code extensions
“It's foolish to think we know who the big names will be 20 years down the line”. Treat today’s market leaders as suppliers to assess, not permanent fixtures.
“A tool that refuses to work 50% of the time is not useful”. Check whether a paid assistant reliably handles your team’s permitted tasks before making it central to the workflow.
↳ Re: Sam Altman’s comments on consumers using RAM
A shared project binder for AI coding tools could reduce the time teams spend re-explaining a job.
What it does: This persistent memory compression system captures an AI worker’s activity, condenses it and brings relevant context into later sessions. Its description lists Claude Code, Codex and Gemini among the supported tools.
Problem it solves: An AI coding session can end before the next session has the background needed to continue the work.
Who should use it: Software teams that use more than one AI coding assistant on continuing projects.
Expected outcomes: A later session can pick up with a record of earlier decisions and work instead of rebuilding that context from scratch.
■ Two Steps Ahead
Over the next 30–90 days, teams using the Claude Code extension in Cursor have two distinct questions for their software providers: who can approve extension updates, and where should review happen before AI-written changes are accepted? The security incident makes the first question urgent; the review workflow makes the second more concrete. Ask for separate answers rather than one broad assurance that AI-written code is checked.
OpenAI’s planned website preference signal gives publishers and other site owners a reason to track changes to their existing settings over the coming months; replacing those settings now would be premature. For restaurants, Zest’s text-message approach is worth watching alongside DoorDash’s reported AI-generated orders: customer-facing AI discovery and AI-completed purchases need different measures of success. Teams considering claude-mem should likewise decide which project details belong in a reusable binder before carrying that history between tools.
Tools & Platforms to Watch
Claude Code: Review AI-written software changes within a development team.
