The Horizon

This week in AI: 8 stories you might have missed (Nov 1–7, 2026)

7 November 2026 · Hossam EL-Kashef

A short weekly recap of the AI and tech stories worth knowing about if you build software. Facts only, with sources for each one.

1. An AI found 948 ways to root Linux, for $1.2M

OpenAI published “A thousand fewer paths to root” on November 4. In one run in mid-October, its pipeline dynamically reproduced 1,724 Linux kernel vulnerabilities and produced 948 working local privilege escalation exploits for a commonly used Linux distribution. The run cost $1,213,430 in API credits, about $1,279 per exploit. These are local exploits: they need an existing low-privileged account. OpenAI’s own advice is that a low-privileged account on a system that isn’t patched should be treated as close to root access for a well-resourced attacker.

Why it matters: finding bugs is getting cheap. Patching fast is the job now.

2. AI agents booked… 4 Uber rides

On its Q3 earnings call on November 4, Uber said third-party AI agents had booked just four rides and about 500 Uber Eats orders in the week before the call. The CEO called it “very, very early”.

Why it matters: agentic shopping gets a lot of headlines. Real usage is close to zero for now.

3. Apple puts engineers on an AI allowance

Bloomberg reported on November 4 that Apple cut some engineers’ daily AI spending caps from $300 to $200, with $100 being discussed, and is steering staff toward cheaper or on-device models.

Why it matters: token spend is becoming a budget line that teams have to justify.

4. Karpathy built a new YouTube front end in about two hours

Andrej Karpathy launched HackerTube on October 31: 6,454 YouTube videos across 150 topics, drawn from about 40,000 Hacker News comments that recommended a video. He writes that Claude did the work, from processing the comments to deploying and hardening the server, and that the whole thing takes about two hours. His manual steps were creating an API key, booting a VPS and buying the domain.

Why it matters: a real, end-to-end example of building by talking to an agent. The idea and the taste were still his.

5. Meta bought an app that runs many coding agents at once

Meta acquired Conductor, an app that runs several coding agents (Claude Code, Codex, Cursor) in parallel, each in its own copy of the codebase. Terms weren’t disclosed. Conductor Cloud goes read-only on November 13 and shuts down on November 20.

Why it matters: parallel agents are where big tech is heading. If your team uses Conductor Cloud, export your work before November 20.

6. Claude Code was reportedly resending whole transcripts for 3–5 word status lines

Users found that Claude Code’s background agents were resending their full transcripts to the model about every 30 seconds, only to produce a short status line. A commenter on Hacker News who said they work on Claude Code wrote that a fix shipped for the most expensive case.

Why it matters: background calls can quietly eat a token budget. Check what your agents actually send.

7. Glasses that type what you silently mouth

Subvocal, a Y Combinator-backed startup, opened US preorders for glasses that use low-power radio waves to sense facial muscle movement and turn silently mouthed words into text. The company claims more than 200 words per minute. The early-bird price is $499, shipping is expected in Q1 2027, and FCC approval is still pending.

Why it matters: a possible new way to give apps and assistants text input. The speed claim hasn’t been independently tested yet.

8. Claude helped crack part of a 40-year-old cell mystery

Anthropic’s biology lab published the first 3D structure of a vault particle found in prokaryotic cells. Vaults were discovered in 1986 and their function is still unknown. The team used Claude to search about 15 billion predicted proteins, and an optimised purification took about 90 minutes instead of the conventional three days. Anthropic is also open about the mistakes: one protocol Claude wrote set the project back two weeks, and it produced a flawed map that the scientists caught.

Why it matters: an honest look at agents in a real lab. Big speed-ups, and people still catching the errors.

Sources