o0o0o0o / 005 30 Oct 2026
Who controls AI?
Models find vulnerabilities, agents take over software development, and companies spend hundreds of billions of dollars. Who sets the limits, who checks the result and who is accountable?
Editorial
A year ago the question was: what can AI do? This week the question changed. An internal Anthropic model chained vulnerabilities into attacks on Chrome, Safari, Android, Windows and Linux on its own. OpenAI said its monitors now watch agents in training, evaluations and internal use, because one incident began exactly where no monitor was running. More than 40% of pull requests on GitHub in September were started by an agent. Capability is no longer in doubt. What is in doubt is who sets the limits.
That is why the theme of issue five is control. Not in a doom-laden sense, but in an engineering one: who sees what an agent does, who approves its steps, who checks the result and who pays when something breaks. Page 5 lists the five questions that run through the whole issue: control of data, behavior, software, infrastructure and knowledge.
In this issue
Contents
- 04EditorialWho holds the switchWhy control is the theme of issue five.
- 05Cover storyWho controls AI?Five questions that tie the issue together.
- 06Cover storyThe question is no longer what AI can doFive stories, five angles on one problem.
- 07Cover storyControl is work, not a switchFive control points and what is open.
- 08SecurityWhen AI answers with someone else's dataCloudflare, GLM-5.3 and 155 responses.
- 09SecurityHow the cache became shared memoryThe KV cache, explained.
- 10SecurityAI no longer just finds bugs. It chains them into attacks.Anthropic and five attack chains.
- 11SecurityFive chains, five vendorsWhat they look like and who patched them.
- 12SecuritySix billion dollars and one chain of compromiseBinance and what is still unknown.
- 13SecurityEstimated, frozen, lost: three different numbersHow to read the amounts.
- 14SignalWho watches agents while they workOpenAI and four layers of monitoring.
- 15SignalTests that show, but do not proveDeepSeek, Qwen and the limits of experiments.
- 16SignalRisk written into the prospectusFour Anthropic models and other people's systems.
- 17SignalFive values instead of a style guideOpenAI's new Model Spec.
- 18Patch nowPatch nowGitLab, Cisco, Collabora, Arista.
- 19Patch nowPHP, Apple and secrets on GitHubPHP2Shell, Brotli, iOS 27.2.
- 20Patch nowPatches of the weekTen CISA KEV additions.
- 21Model watchGemini 4 Carbon: a model for cyber defenseLimited access, vendor figures.
- 22Model watchTerra is a model for work, not for chatGPT-6.1 Terra and where to find it.
- 23Model watchModels of the weekCarbon, Terra and Kimi K3.1 Pro.
- 24BuildProgramming has changed shapeGitHub Universe 2026: GA and preview.
- 25BuildWho writes the first line of code now?Octoverse: agents start 40% of PRs.
- 26BuildA record year in numbersOctoverse 2026 in six charts.
- 27BuildFusion picks the model for youCopilot Fusion is generally available.
- 28BuildAn agent that works when you do notCloud, Agent Merge, permissions.
- 29BuildAgents in Actions, scanners in the repositoryAgentic Workflows and AI Scan.
- 30BuildGitLab Orbit gives agents a map of the projectThe context graph is GA.
- 31BuildAndroid Studio lets your agents into the IDERabbit 2 and ACP.
- 32BuildDuckDB 2.0 knows when an agent is drivingLTS, Quack and VARIANT.
- 33BuildAgent skills get a shared homeAgent Skills in a foundation.
- 34Use thisLock your actions in GitHub ActionsA lockfile in six steps.
- 35MoneyAI brings in revenue. But what does growth cost?Microsoft, Alphabet, Amazon.
- 36MoneyThree reports on one pageSix charts, one quarter.
- 37MoneyAnthropic opens its books before going publicS-1: revenue, loss, offering.
- 38MoneyWhat the prospectus says about risk and controlTrust, partners, open questions.
- 39InfraThe debt building data centersMeta, SpaceX, Broadcom and SPVs.
- 40InfraMemory is the new bottleneckSK hynix and Intel.
- 41HardwareMacBook Pro gets a touchscreenPROLED XDR, M5 Pro and M5 Max.
- 42HardwareiMac M6, XREAL Aura and iOS 27.2Three devices and one update.
- 43LabAI wrote a proof. What do mathematicians do now?Three OpenAI results.
- 44LabThree results, three ways to checkWhat is public, who checks.
- 45LabThree problems, explained without formulasKakeya, ETH, Yang-Mills.
- 46LabGenie 4: a world that forms as you watchDeepMind's research preview.
- 47LabHumanoids grow faster than profitsUnitree after its listing.
- 48OrbitOne light-day from homeVoyager 1 and 18 November.
- 49OrbitHow far is one light-dayScale, CRISTAL and MMX.
- 50WorkAgents enter the officeServiceNow, Signals, Microsoft.
- 51ShippedShipped this weekWhat is actually available.
- 525 minutesIn briefEight stories in five minutes.
- 53NumbersThe week in numbersSecurity, models, code.
- 54NumbersMore numbers of the weekMoney, memory, space.
- 55NextWhat we are watchingDates ahead.
- 56SourcesAll sources in this issuePrimary sources first.
