AI Product Updates Daily — June 7, 2026

The day before Apple's biggest keynote in years, the AI world kept moving: Anthropic published hard numbers on how fast Claude is rewriting the rules of software development, OpenAI shipped a security feature designed for organizations handling sensitive data, Google confirmed its engineers are riding the same AI-coding wave, a new Microsoft desktop AI agent went public, and Cloudflare data showed that bots just quietly overtook humans on the open web.

Anthropic: Claude now writes over 80% of its own code — and calls for a global slowdown

Anthropic published a detailed research post titled "When AI Builds Itself," measuring Claude's progress toward recursive self-improvement — the threshold where an AI can fully design its own successors without human intervention. The company says it is not there yet, but the internal metrics are striking. 1

As of May 2026, more than 80% of the code merged into Anthropic's production codebase was authored by Claude. Anthropic engineers now merge 8× as much code per quarter as the 2021–2025 pre-Claude baseline. Before Claude Code's research preview launched in February 2025, AI-authored code was in the low single digits.

On benchmark performance, the length of tasks Claude can complete reliably has been doubling roughly every four months:

March 2024 (Claude Opus 3): 4-minute software tasks
March 2025 (Claude Sonnet 3.7): 1.5-hour software tasks
March 2026 (Claude Opus 4.6): 12-hour software tasks
Current (Claude Mythos Preview): at least 16 hours, at the upper limit of METR's measurement capability

Line graph showing Claude Code session success rate across task complexity tiers from September 2025 to May 2026 — Claude Code session success rates across task complexity tiers, September 2025–May 2026 1

On research tasks, Claude Mythos Preview beat human researchers' next-step choices 64% of the time in 129 evaluated sessions — up from 22% for Claude Haiku 3 in March 2024. An April 2026 open-ended AI safety experiment found Claude agents closed 97% of a targeted performance gap in 800 cumulative hours at $18,000 compute cost; a team of human researchers closed 23% of the same gap in one week.

Alongside these figures, Anthropic is calling for a coordinated global pause in AI development. Co-founder Jack Clark placed the probability of state-of-the-art AI achieving recursive self-improvement by late 2028 at roughly 60%.

anthropic.comhttps://www.anthropic.com/institute/recursive-self-improvement外部リンク

コンテンツカードを読み込んでいます…

OpenAI: ChatGPT gets Lockdown Mode and an upgraded memory system

Lockdown Mode is OpenAI's new security feature for ChatGPT, aimed at organizations and individuals handling sensitive data. When turned on, it disables live web browsing (cached content only), image retrieval from the web, deep research, and agent mode. 2

The goal is to reduce the chance that sensitive data gets exfiltrated through prompt injection — attacks where malicious instructions are hidden inside webpages, uploaded files, or other content ChatGPT reads. OpenAI is explicit that the mode does not eliminate injection risk entirely: a cached page or uploaded document could still influence responses. Rollout is to self-serve ChatGPT Business accounts and eligible personal accounts.

Separately, OpenAI rolled out an upgraded memory architecture — internally called an improved "dreaming" system — to ChatGPT Plus and Pro users in the US. 3 The update targets three common failure modes: carrying forward stale context, missing user preferences, and scaling memory to hundreds of millions of accounts. Users can review a memory summary page, correct entries, and delete stored details. Free and international rollout is planned for coming weeks.

Google: 75% of new code is AI-generated, publisher opt-out goes live June 17

CEO Sundar Pichai disclosed that 75% of Google's new internal code is now AI-generated, reviewed by engineers before it merges. 4 The figure has moved fast: roughly 25% in October 2024, 50% in autumn 2025, 75% by April 2026. Pichai said a recent AI-and-engineer collaboration on a complex code migration ran six times faster than a comparable human-only effort a year earlier.

On search, Google will start honoring its new AI Overviews and AI Mode opt-out toggle on June 17, 2026. 5 The control, deployed in Google Search Console for a subset of UK sites, lets publishers remove their content from AI Overviews, AI Mode, and Discover's AI features while keeping normal organic rankings. The UK Competition and Markets Authority's binding conduct requirement — the world's first — takes legal effect December 3, 2026. Google says the opt-out is not a ranking signal and does not affect the Gemini app.

Microsoft: Scout desktop AI agent enters public preview

Microsoft released Microsoft Scout (Frontier) as an early public preview on June 6, available for download on Windows 11 and macOS. 6 Scout is a desktop AI agent that works across local files, browser tabs, and Microsoft 365 surfaces. Its stated capability range includes meeting prep, workflow automation, browser-based task execution, and developer tooling (optional).

All outbound actions — sending emails, posting to Teams, updating calendar events, running privileged operations — require explicit user approval before execution. Access requires enrollment in Microsoft's Frontier program, admin enablement, a GitHub Copilot Business or Enterprise subscription, and Intune-managed devices.

The release arrives shortly after Microsoft's Build 2026 announcement of seven in-house MAI models, including MAI-Thinking-1 (its first in-house reasoning model), MAI-Code-1-Flash (a 5-billion-parameter coding model for GitHub Copilot), and MAI-Image-2.5. 7

Cloudflare: Bots now generate more web traffic than humans

Cloudflare CEO Matthew Prince confirmed that automated bots and AI agents now account for 57.3% of global web traffic — the first time in internet history that non-human traffic has surpassed human browsing. 8 The data covers the week of May 29–June 4, 2026. Prince had projected this crossover would happen in 2027.

The acceleration is largely driven by AI agents that scan dozens or hundreds of pages to answer a single query, generating far more HTTP requests per task than a human searcher would. The trend is creating fresh pressure on advertising-supported publishers, since bots do not view ads, subscribe, or otherwise monetize like human readers.

Apple WWDC26 keynote: Monday June 8, 10 a.m. PT

Apple's Worldwide Developers Conference keynote is tomorrow. 9 The headlining expectation is a heavily rebuilt Siri — internally codenamed Campo — built on a conversational AI architecture, reportedly powered by Google Gemini and supported by Nvidia compute. Speculation points to a standalone Siri app and a Dynamic Island presence that illuminates when Siri is active (Apple's WWDC promo tagline is "All systems glow").

iOS 27 is expected to be a refinement release, modeled on Mac OS X Snow Leopard: code cleanup, performance gains, and bug fixes in response to complaints about iOS 26's battery life, Liquid Glass UI glitches, and keyboard issues. New AI features include AI photo editing, AI text generation tools, and an option to choose a third-party AI model for Apple Intelligence features. Preparations for a foldable iPhone are also embedded in the release, with a reported "Parallel View" landscape mode for running two apps side by side.

WWDC26 promo artwork — Apple's tagline "All systems glow" points to a revamped Siri 9

Quick hits

Google / SpaceX: Google will pay SpaceX $920 million per month from October 2026 through June 2029 for access to roughly 110,000 NVIDIA GPUs, per a SpaceX regulatory filing. Google framed the deal as bridge capacity for its Gemini Enterprise agent platform. 10
White House AI advisor: Sriram Krishnan is leaving his role as the White House AI policy advisor, per TechCrunch. 11
Trump / OpenAI: The Trump administration is reportedly in talks to take an equity stake in OpenAI, per TechCrunch. 12