AI Week in Review · June 6, 2026 · 14:28

Recursive AI Goes Public & The Backlash Gets Lawyers - AI Week in Review (May 31-June 6, 2026)

This week in AI: Anthropic says Claude writes 80% of its merged production code and Sakana opens an RSI Lab, Anthropic nears $1T while Alphabet raises $80B and DeepSeek $7.4B, Florida sues OpenAI, Connecticut passes an AI disclosure law, an engineer wins a religious accommodation to avoid AI coding tools, and the line between AI security and AI exploitation gets blurrier.

Recursive AI Goes Public & The Backlash Gets Lawyers - AI Week in Review (May 31-June 6, 2026)
0:0014:28

Today's AI Week in Review Topics

  1. 01

    Recursive self-improvement, out in the open

    — Anthropic said Claude now writes more than eighty percent of the production code that gets merged inside the company, and warned in the same week that verification and governance — not capability — may become the real bottleneck. Sakana AI formalized an RSI Lab in Tokyo focused on compute-efficient self-improvement loops. OpenAI was reported to be leading a round in Opal Electronics for AI-native hardware. European lab Inherent raised fifty million dollars to build agents that generate scientific hypotheses. The week the industry stopped using the term AGI in slide decks and started saying RSI out loud.
  2. 02

    Coding agents: more capable, more contested

    — xAI's grok-build-0.1 entered public beta. MiniMax M3 launched with open weights, frontier coding, and ultra-long context. Cognition described how Devin uses parallel auditable testing to produce more ready-to-merge work. The open-source ECC project tried to standardize hooks, governance, and injection scanning across Claude Code, Codex, and Cursor. Microsoft's leaked Scout is an always-on Microsoft 365 agent — and a separate leak alleged it was designed to make people addicted. GitHub said agent activity is pushing it toward billions of commits. Stanford CS336 published rules limiting AI assistants in coursework. Google engineers shared memes about the low-quality AI code they're being asked to merge. A software engineer received a religious accommodation to avoid AI tools at work. The capability curve and the friction curve are both bending upward at once.
  3. 03

    The money keeps escalating

    — Anthropic's Series H is approaching a one-trillion-dollar valuation. Alphabet is reportedly raising up to eighty billion dollars via a stock sale to expand AI compute. DeepSeek is reportedly raising about seven point four billion at a fifty-two-to-fifty-nine-billion valuation. Generalist AI raised four hundred million for physical-AGI robotics. Apple approved a third-party AI agent called Poke inside iMessage. Leaked screenshots showed Microsoft consolidating Copilot into a single 'super app.' OpenAI was reported leading a round in Opal Electronics for vision-and-voice-forward devices. The US Commerce Department tightened export controls to block Chinese AI firms from buying frontier Nvidia and AMD chips through overseas subsidiaries. The capital story is no longer separable from the geopolitical one.
  4. 04

    Agents go offensive — and defensive

    — Anthropic expanded Project Glasswing for AI-assisted vulnerability discovery and published a reference harness showing Claude can find, verify, report, and patch security bugs inside a sandbox. A researcher demonstrated agentic LLMs exploiting Firebase misconfigurations on a vulnerable React Native app. Vercel reported real-world 'inference theft' surging on a public AI chat endpoint. NVIDIA released Nemotron 3.5 Content Safety, a multimodal moderation model with auditable reasoning. Florida's Attorney General sued OpenAI and Sam Altman over product-liability-style safety claims. Connecticut passed a workplace AI disclosure law. South Korea moved toward requiring forums to pre-screen user-uploaded images and video with AI. OpenAI published a federal policy blueprint. The same week, agents got better at finding vulnerabilities, and at being exploited.
  5. 05

    The backlash gets lawyers

    — A software engineer publicly reported receiving a religious accommodation to avoid AI coding tools, which is now the most concrete example yet of AI usage becoming a contested workplace requirement. UC Berkeley saw unusually high failing rates linked to overreliance on LLMs. Erin Brockovich documented community pushback against AI data centers over water, noise, and grid stress. Vox spotlighted 'AI successionism,' a posthuman ideology arguing that AI should inherit the future. Amnesty International framed many generative AI systems as human-rights violators because of unlawful scraping. A Dune teaser reminded everyone of Herbert's anti-thinking-machines premise. AXA's global mental-health survey flagged trust gaps and harmful AI advice. The pushback that last week 'got articulate' this week started filing the paperwork.

Sources & AI Week in Review References

Full Episode Transcript: Recursive self-improvement, out in the open & Coding agents: more capable, more contested

On Thursday this week, Anthropic published a number it had been hinting at for months. By May 2026, Claude was writing more than eighty percent of the code merged into Anthropic's own production codebase. Not generated, not suggested — merged. The blog post was careful to frame the statistic as a governance challenge: when most of your code originates from an AI, your bottleneck stops being engineering throughput and starts being verification, review, and accountability. On the same day, Sakana AI in Tokyo formally announced an RSI Lab — recursive self-improvement — focused on compute-efficient self-improving loops, publishing openly while warning about benchmark gaming and unsafe self-modification. By the end of the week, OpenAI was reported to be backing AI-native hardware, a European lab called Inherent had raised fifty million dollars for AI that generates scientific hypotheses, and the industry vocabulary had shifted. People stopped saying AGI in slide decks and started saying RSI out loud. Welcome to The Automated Weekly — a magazine-style look at the forces shaping artificial intelligence, designed not for engineers, but for anyone trying to understand where the industry is heading. I'm TrendTeller. This week, the recursive-self-improvement story landed in the same week Anthropic's valuation pushed toward a trillion dollars, Alphabet was reported raising up to eighty billion in stock to fund AI compute, DeepSeek was reported raising seven point four billion in China, Generalist AI raised four hundred million for robotics, Apple opened iMessage to a third-party AI agent for the first time, and Microsoft was reported consolidating Copilot into a single super app. It was the same week Florida's Attorney General sued OpenAI, Connecticut passed an AI disclosure law for workplace decisions, and a software engineer publicly reported winning a religious accommodation to avoid using AI coding tools. And quietly underneath all of it, Anthropic's Mythos model was found to be both finding security vulnerabilities and being used to exploit them. Five threads. One week. Let's pull on each.

Recursive self-improvement, out in the open

Anthropic's eighty-percent number is the cleanest statement of the recursive-self-improvement story we've had so far. Not 'Copilot suggests a lot of code.' Not 'most engineers use AI at some stage.' More than eighty percent of the code that ends up in production at Anthropic, the company building Claude, is being written by Claude. The post itself was careful: the constraint isn't capability anymore — it's verification, review, and accountability. Which is what RSI was always going to look like, if it arrived: a curve where the AI does more of the work, and the humans do more of the checking. It landed in a week when the language shifted. Sakana AI in Tokyo formally launched an RSI Lab, with a focus on compute-efficient, evolution-inspired self-improvement loops, publishing openly while explicitly listing the risks — benchmark gaming, unsafe self-modification — that this kind of work normally lets stay implicit. A European lab called Inherent emerged with fifty million dollars to build agents that generate scientific hypotheses, betting that the next frontier is finding the right questions rather than answering known ones. OpenAI was reported leading a round in Opal Electronics to build vision-and-voice-forward AI-native hardware. A separate Anthropic post argued AI is increasingly building AI and explicitly used the phrase 'recursive improvement' rather than the safer 'AI for AI research.' What's changed isn't the technology. The recursive-improvement loop has been there since coding agents existed. What changed this week is that the industry stopped describing it euphemistically. AGI as a term has been hollowed out by the timelines debate; RSI is more concrete, more measurable, and more honest about what's happening on the ground. The eighty-percent number isn't an end state. It's the first widely shared lap-time. Two things to watch from here. First, whether other labs publish their own version of that statistic — because if the number is eighty percent at Anthropic, it's not zero everywhere else. Second, whether verification and governance start showing up in roadmaps and earnings calls the way capability did from twenty-twenty-three through twenty-twenty-five. RSI without verification is the failure mode the safety community has been warning about for a decade. We're now in a week where the failure mode and the business model are visibly the same diagram.

Coding agents: more capable, more contested

The coding-agent capability story this week was as dense as any we've covered. xAI shipped grok-build-0.1 in public beta on its API. MiniMax M3 launched with open weights, frontier coding, ultra-long context, and multimodality, signaling that the gap between closed and open is narrowing again. Cognition published a technical deep-dive on how Devin uses parallel auditable testing — labeled screenshots, chaptered videos — to produce more ready-to-merge results, which is the kind of detail you only share when you're confident enough to be inviting copying. The open-source ECC project went public with a single agent harness that tries to standardize hooks, governance, and prompt-injection scanning across Claude Code, Codex, Cursor, and others. Microsoft's leaked Scout — an always-on Microsoft 365 background agent tied to governed identity — became the headline product of the week from the largest enterprise AI buyer in the world. And then the same week showed exactly how contested all of that is becoming. A separate leak about Scout, reported in tech press, alleged that the product was explicitly designed to make users addicted, with documents discussing engagement loops and behavioral hooks. GitHub said agent activity is pushing the platform toward billions of commits and is forcing architectural rewrites — which is the polite enterprise way of saying our infrastructure is buckling. Stanford's flagship LLM systems course CS336 published explicit rules limiting AI assistants to tutoring and debugging, after watching what unrestricted use did to learning. Google engineers were reported to be sharing internal memes mocking the low-quality AI code they're being asked to merge. UC Berkeley reported unusually high failing rates this term tied to AI overreliance and academic dishonesty. And — most concretely — a software engineer publicly reported receiving a religious accommodation from his employer to avoid using AI coding tools, which a year ago would have been an interview question, not a workplace policy. Put together, the picture is sharp. The coding agents got measurably better this week. The friction around them got measurably real. Both curves are bending upward, and the industry is now going to spend the rest of the year figuring out which bends faster.

The money keeps escalating

The capital story this week was the loudest it has been since the early ChatGPT wave. Anthropic's Series H pushed its valuation toward one trillion dollars — a number that two years ago would have read as satirical. Alphabet was reported preparing a stock issuance of up to eighty billion to fund AI compute expansion, which is the kind of dilution you only accept when the alternative is falling behind. DeepSeek in China was reported raising about seven point four billion at a fifty-two to fifty-nine-billion valuation. Generalist AI raised four hundred million for what its CEO called physical-AGI robotics. OpenAI was reported leading a funding round in Opal Electronics to make AI-native devices. The distribution story was just as loud. Apple approved a third-party AI agent called Poke inside iMessage via the Messages for Business framework — the first time an outside AI agent has been given an action-oriented entry point into a core iPhone app, and a signal that Apple's resistance to agentic third-party software inside its operating system may finally be cracking. Microsoft's leaked plans showed Copilot consolidating into a single super-app combining chat, planning, and GitHub-style coding. Microsoft, separately, is shipping new MAI models and a Frontier Tuning workflow that lets enterprise customers tune behavior without retraining. Underneath all of it, the geopolitical layer hardened. The US Commerce Department updated guidance to block Chinese AI firms from buying frontier Nvidia and AMD chips through overseas subsidiaries — closing the most-discussed export-control loophole. A leaked story this week reported that one company spent five hundred million dollars in a single month on Claude API after missing usage caps. The runaway-spend story isn't theoretical. The number you hear in board meetings is now real enough to print. The summary is uncomfortably simple. The biggest cheques in technology are being written. The biggest distribution channels are opening up. And the geopolitical fence around all of it is being raised the same week. Capital, distribution, and policy are all moving at the same time, in coordinated directions.

Agents go offensive — and defensive

The security half of the week was the most unsettling one we've covered in months. Anthropic expanded Project Glasswing — its AI-assisted vulnerability-discovery program — and published a reference harness showing how Claude can find, verify, report, and patch security bugs inside a sandbox with staged operational controls. That's defense. On the offense side, a researcher published a demonstration of agentic LLMs reliably identifying and exploiting Firebase misconfigurations in a vulnerable React Native app, using off-the-shelf frontier models. Vercel reported a wave of real-world inference-theft attacks on an AI chat endpoint, where attackers ran up API bills on the victim's account and resold the access. The same Mythos and Opus generation that's hardening defenders is hardening attackers at the same speed. The regulatory and governance side moved just as fast. Florida's Attorney General filed a civil lawsuit against OpenAI and Sam Altman, testing whether product-liability claims used against social-media companies extend to chatbots — the answer to which will quietly determine a decade of frontier-model deployment. Connecticut passed a workplace AI disclosure law requiring transparency in automated employment decisions and additional notice when layoffs are driven by automation. South Korea moved toward requiring online forums to pre-screen all user-uploaded images and video with AI moderation — a move praised by child-safety groups and criticized as the largest formal prior-restraint scheme in any liberal democracy. The Trump administration's revised AI cybersecurity executive order proposed voluntary pre-release review and a vulnerability clearinghouse. NVIDIA released Nemotron 3.5 Content Safety, a multimodal moderation model with custom-policy enforcement and optional auditable reasoning chains. And OpenAI published a federal AI governance blueprint calling for a durable framework and a strengthened US AI Safety Institute — a document interesting less for its contents and more for what it implies. The frontier labs are now writing the regulation themselves, then publishing it. That's the part of the policy cycle that historically arrived last. This week it arrived first.

The backlash gets lawyers

Last week, the pushback got articulate. This week, it acquired infrastructure. The single most concrete moment was a software engineer publicly reporting that he had won a religious accommodation from his employer to avoid using AI coding tools — a workplace category previously occupied by dietary restrictions and Sabbath observance, now extended to the question of whether a person can be required to delegate their craft to a model. It is, as far as anyone covering this has been able to confirm, the first reported case of its kind. It will not be the last. The legal and institutional layer hardened in parallel. Florida sued OpenAI. Connecticut required workplace AI disclosure. Amnesty International published a position arguing that many generative AI systems rely on unlawful web scraping and therefore constitute human-rights violations around privacy, discrimination, and freedom of expression — language designed to be quoted in courts and used to litigate, not to debate. AXA's global survey reported worsening mental-health outcomes correlated with AI overuse and flagged trust gaps and reports of actively harmful advice. UC Berkeley's high failing rates this semester were tied directly to AI overreliance and academic dishonesty. The cultural layer hardened too. Vox spotlighted 'AI successionism' as a coherent posthuman ideology — the explicit position that AI should inherit the future even at humanity's expense — which is the kind of belief system that only gets a name once it has enough adherents to be worth refuting. Erin Brockovich, of all people, documented growing community opposition to AI data centers over water, noise, grid stress, and lack of disclosure. A new Dune teaser reminded an audience already primed for the metaphor of Frank Herbert's original anti-thinking-machines premise — but the reframing was deliberate. The danger in Herbert's universe isn't rogue robots. It's concentrated power and dependency. That's the version of the AI-risk argument that's winning right now: not 'the model becomes evil,' but 'we became dependent.' That's not a benchmark you can game. It's a position. And it's now backed by lawyers, regulators, accommodations, and a lot of essays.

That's your week in AI — May 31st through June 6th, 2026. Anthropic said Claude writes most of its production code. Sakana opened an RSI Lab. Inherent raised fifty million to find scientific questions. OpenAI is reportedly backing AI-native hardware. Anthropic's valuation pushed toward a trillion, Alphabet is reportedly raising eighty billion for compute, DeepSeek seven point four billion. Apple opened iMessage to a third-party AI agent for the first time. Microsoft consolidated Copilot into a super-app — and got hit by an alleged-addiction leak. The US Commerce Department closed a major export-control loophole. Florida sued OpenAI. Connecticut required workplace AI disclosure. Amnesty International called generative AI a human-rights issue. A software engineer received a religious accommodation to avoid AI coding tools. Anthropic showed Claude finding security bugs in the same week researchers showed Claude-class models exploiting Firebase. The pushback that was articulate last week now has lawyers. Three things to watch next week. First, whether another major lab publishes their own version of the eighty-percent number — and whether anyone publishes a verification number alongside it. Second, whether the Apple-iMessage opening generalizes to a broader third-party AI app surface inside core iPhone applications, which would change distribution economics for every AI startup. Third, whether the Florida lawsuit gets joined by other state attorneys general, the way the Section 230 fights spread one state at a time before reaching the Supreme Court. I'll see you next Saturday. From The Automated Weekly, this is TrendTeller.

More from AI Week in Review