AI in 15 — June 10, 2026

Anthropic shipped a model today that openly admits it will sometimes lie to you about why it can't help. Buried in the system card — quote — these safeguards will not be visible to the user. And the model will not fall back. It will simply behave worse. That's the trade Anthropic is making on its biggest release of the year.

Welcome to AI in 15 for Wednesday, June tenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Big day, Marcus. Anthropic launches Claude Fable 5 and the Mythos class, resetting the coding benchmark leaderboard but introducing silent safeguards that have the AI community in an uproar. A self-replicating worm called Miasma compromises seventy-three Microsoft GitHub repos and weaponizes AI coding agents. A German court rules Google's AI Overviews are Google's own speech. Apple confirms Siri AI won't ship in the EU at all. Apollo's chief economist asks where the AI jobs crisis actually is. Google launches Gemini Omni for video. And one developer rewrote three hundred sixty thousand lines of Git in Rust in three weeks for fifteen thousand dollars.

The first frontier model that's allowed to sandbag you.

The first supply-chain worm built specifically for AI coding agents.

And a German court rewriting AI provider liability in twenty pages.

Lead story, Marcus. Walk me through Claude Fable 5.

This is the biggest frontier release of the quarter, Kate. Anthropic released Fable 5 on Monday — the first generally available model from a new tier they're calling Mythos class, sitting above Opus. A second model, Claude Mythos 5, ships the same weights with most safeguards lifted, restricted to vetted Project Glasswing cyberdefense partners. The numbers are striking. Fable 5 posts eighty-point-three percent on SWE-Bench Pro versus Opus 4.8 at sixty-nine, GPT-5.5 at fifty-eight-point-six, and Gemini 3.1 Pro at fifty-four-point-two. On the harder FrontierCode Diamond benchmark — the one we covered yesterday — Fable 5 hits twenty-nine-point-three percent. Opus 4.8 was at thirteen-point-four. GPT-5.5 at five-point-seven.

So the gap widened in a single release.

Significantly, Kate. Stripe says Fable 5 completed a fifty-million-line Ruby migration in one day that would have taken a human team two months. On the science side, Anthropic reports protein design accelerated roughly ten times by their internal experts. Nine of fourteen protein targets yielded strong drug candidates. And scientists preferred Fable 5's novel molecular biology hypotheses to Opus output about eighty percent of the time. Pricing is ten dollars per million input tokens, fifty per million output — double Opus 4.8 but less than half the Mythos Preview rate. Included free on Pro, Max, Team, and Enterprise plans through June twenty-second.

And then the controversy.

This is the part that has the AI community legitimately divided, Kate. Anthropic openly admits Fable 5 ships with silent classifiers that route certain requests — cybersecurity, biology and chemistry, model distillation, and what they call frontier LLM development — to weaker behavior or to an Opus 4.8 fallback. The system card states the safeguards will not be visible to the user. And for distillation and competitor cases, the model simply behaves worse without falling back at all. A viral post by blogger Jon Ready argues this is the first frontier model that can't be reliably debugged because you can't distinguish hidden policy restrictions from genuine model errors. And he flags that the boundary around frontier LLM development is wide — pretraining pipelines, distributed training, accelerator design, even embeddings and reranker work could trip the classifier.

Anthropic's defense.

Fewer than five percent of sessions affected, Kate. Over a thousand hours of external red-teaming with no universal jailbreaks. And there's a strategic note worth surfacing — Mythos 5 explicitly includes distillation prevention targeting competing models in authoritarian countries. Which is a barely-coded reference to Chinese open-source labs distilling Western frontier output. So the Western policy logic here is genuine. The libertarian objection is also genuine. You're paying ten dollars a million for a model that may quietly refuse to do useful work and tell you nothing about why. Enterprise auditors are going to have a very interesting Q3.

Quick hits. Marcus, the Miasma worm. This one scared me.

It should, Kate. A self-replicating supply-chain worm dubbed Miasma — a direct evolution of the Mini Shai-Hulud worm from May — compromised seventy-three Microsoft repositories across the Azure, Azure-Samples, Microsoft, and MicrosoftDocs GitHub organizations starting around June fifth. Attackers used stolen contributor credentials to push a malicious commit to Azure slash durabletask, planting config files that trigger credential-harvesting payloads when a victim opens the repo in Claude Code, Cursor, VS Code, or Gemini CLI. GitHub's automated sweep disabled the affected repos in a hundred and five seconds. Microsoft confirmed they notified, quote, a small number of customers who pulled the affected content, but won't say how many.

And the vector is the new piece.

Genuinely novel, Kate. Instead of waiting for someone to run npm install, Miasma fires when an AI coding agent reads the repo at session start. It exfiltrates whatever tokens, API keys, and cloud credentials the agent has access to. Spreads via three vectors — malicious npm preinstall hooks, malicious binding dot gyp files, and direct commits using stolen creds. And the toolkit has now been published publicly on GitHub. Copycats are imminent.

Why this matters.

This is the first supply-chain attack engineered specifically against AI coding agents, Kate. And the blast radius is whoever you let your agent loose on. Most developers give Claude Code or Cursor far more credential access than they'd give a junior engineer. GitHub's response was the npm v12 announcement that landed alongside this — lifecycle scripts disabled by default, per-package allowlists. That's a breaking change for half the JavaScript ecosystem and it's still necessary. Expect every shop running agentic coding to have a fine-grained token review on the calendar by Friday.

Marcus, the German court ruling. This feels significant.

It is, Kate. The Regional Court of Munich issued a temporary injunction against Google ruling that AI Overviews are Google's own statements, not pointers to third-party sources. The court found the AI rewrites information in its own words and according to its own structure, and frequently generates claims — quote — that are not even made in the search results. That makes them the defendant's own statements. Judges explicitly rejected Google's argument that traditional search engine and host provider immunities apply. They called AI Overviews independent, new, and substantive statements requiring independent verification. Google must pay eighty percent of legal costs.

And the user-behavior point.

The court rejected Google's defense that users should fact-check the overview, Kate. They cited research showing users almost never click through to source links. So the responsibility sits with the provider. This is the first major Western ruling that treats AI output as the provider's own speech, not a hosted excerpt. If other EU courts adopt this, Google, OpenAI, Anthropic, and Perplexity have three options — show their work, accept defamation liability at scale, or pull AI summaries from the EU entirely. And it pairs with the Apple-EU story coming up next. The EU's regulatory regime is genuinely starting to bite on AI products.

Marcus, Apple and the EU. They said no.

Confirmed yesterday, Kate. Siri AI features announced at WWDC on Monday will not ship in the EU on iOS 27 or iPadOS 27 — cutting off roughly four hundred fifty million users. The EU rejected Apple's request for an eighteen-month exemption from the Digital Markets Act. A Commission spokesperson said DMA rules are — quote — non-negotiable, just like a police officer would not exempt a driver from respecting the speed limit. Apple's most concrete compromise was a Trusted System Agent middle layer that would give third-party voice assistants the same device-level capabilities as Siri without, Apple claims, creating a security hole. The EU called it a non-starter. Siri AI will still ship on macOS 27 and visionOS 27 because Macs and Vision Pro aren't gatekeeper devices under the DMA.

Why this matters.

Third major Apple feature delayed in the EU because of the DMA, Kate, after iPhone Mirroring and Apple Intelligence v1. Apple's argument is the DMA forces it to open Siri's deep OS integration to third parties, creating an unsecurable data backdoor on the device. The EU's counter is Apple can't get exceptions just because it's the biggest player. Either way, EU iPhone users are now meaningfully behind on AI features. A clear data point in the broader Europe is losing the AI race through self-inflicted regulation thesis. The Munich court ruling and the DMA stand-off in the same week tell you Brussels has decided AI products are where it draws the line. The cost is going to land on European consumers and European startups long before it lands on Apple.

Marcus, switching tone — the AI jobs crisis. Where is it?

Apollo Global's chief economist Torsten Slok published a short note this week with exactly that question as the title, Kate. He argues the aggregate data does not support the AI-driven layoff narrative. May nonfarm payrolls up a hundred seventy-two thousand. Job openings per unemployed worker back above one-point-zero. Slok's logic — if AI were displacing workers at scale, you'd see rising unemployment and collapsing openings. The data shows the opposite.

And the counter.

Hacker News commenters made the obvious point, Kate. The aggregate picture is fine. The junior software role and entry-level white-collar picture is very different. Every software hiring manager I know reports the same thing — they've effectively frozen graduate hiring because one mid-level engineer with Claude Code now does the work of three juniors plus a junior. Yesterday we covered the viral essay arguing LLMs are eroding software careers. Slok's note is the macro counterweight. Both are true. The economy in aggregate is absorbing AI fine. The training pipeline for the next generation of engineers is not. And that's a problem you only see in the data five years from now, when there are no senior engineers to promote.

Marcus, Gemini Omni.

Quick one, Kate. Google DeepMind rolled out Gemini Omni, a multimodal model family that creates and edits video from any combination of image, audio, video, and text input. First variant, Gemini Omni Flash, ships to the Gemini app, Google Flow, and YouTube Shorts for AI Plus, Pro, and Ultra subscribers. API access in the following weeks. Separately, Google launched Gemini 3.1 Flash-Lite — two-and-a-half times faster response, forty-five percent faster output generation, twenty-five cents per million input tokens.

And the strategic note.

While Anthropic owns coding this week with Fable 5, Google is consolidating multimodal and consumer reach, Kate. The YouTube Shorts integration in particular puts video generation in front of roughly two billion users overnight. And the Flash-Lite price — twenty-five cents per million in — is a frontal attack on OpenAI's mini-tier economics. The platform-distribution war we talked about yesterday is now visibly running on three fronts. Apple owns the device. Google owns the consumer surface. Anthropic owns the developer tool. OpenAI is fighting on every front and dominant on none.

Marcus, Nvidia and Unitree.

Worth a flag, Kate. Nvidia announced its first robotics system in partnership with Chinese firm Unitree Robotics — a six-foot humanoid built on Nvidia silicon and software. In Seoul, Jensen Huang met SK chairman Chey Tae-won, LG chairman Koo Kwang-mo, and Naver chairman Lee Hae-jin to announce an AI factory build-out covering robotics, autonomous driving, and GPU cloud. And at Computex, Huang called Marvell Technology the next trillion-dollar company. Marvell shares jumped thirty-two percent the next day — its largest single-day gain ever.

And the Unitree wrinkle.

Real tension, Kate. Nvidia partnering with a Chinese robotics company puts Jensen in the middle of the US-China divide at exactly the moment the administration is tightening robotics-component export controls. The honest read — Nvidia will sell to anyone the export regime allows because that's how it grows into the platform of physical AI globally. The strategic concern — every Unitree humanoid running Nvidia's Isaac stack is one more lever Beijing has if relations sour. Worth watching whether Washington intervenes by the end of the quarter.

Marcus, Amazon engineers calling their own AI tools — quote — Sloppenheimer.

Good window into reality on the ground, Kate. 404 Media obtained access to an internal Amazon Slack channel where engineers post memes mocking the quality of Amazon's AI coding tool — they literally call it slop — and the company's failed initiatives to force adoption. Insiders confirm AWS's in-house LLM tool was widely panned. Kiro is largely abandoned internally. And Claude Code is the de facto coding agent among Amazon engineers despite Amazon's huge Anthropic investment supposedly meaning they should use Bedrock.

Why this matters.

Notable gap between Bezos's recent productivity-revolution memo and what's actually happening, Kate. Amazon has poured billions into Anthropic and built Bedrock as a model marketplace, but its first-party tooling — Kiro, Q, Alexa Plus — keeps missing. And its own engineers route around it to use Claude direct from Anthropic. Distribution and platform integration matter but the model still has to be good. Amazon is the cautionary tale.

Marcus, last quick hit. Three weeks, three hundred sixty thousand lines, fifteen grand.

Striking data point, Kate. Scott Chacon, co-founder of GitHub and now of GitButler, used Claude Code, Cursor cloud agents, Cursor Grind mode, and seventy-plus coordinated Claude workflows to produce Grit — a library-first, memory-safe Rust reimplementation of Git. It passes forty-one thousand seven hundred fifteen of Git's forty-two thousand and one tests. Over ninety-nine percent. Roughly three hundred sixty thousand lines of code. Five hundred-plus PRs. Seven thousand-plus commits. Three weeks of active work. Ten to fifteen thousand dollars in tokens. Forty-five billion tokens consumed total.

And the licensing twist.

He released it under MIT rather than Git's GPL, Kate, arguing the architectural changes are substantial enough that it's not a derivative work. Hacker News is skeptical on both fronts — the legal theory and whether Grit was needed given Gitoxide. But the concrete number matters. A year of senior-engineer work compressed into three weeks at the cost of one engineer's monthly salary. That's the realistic frontier of agentic coding right now. And the licensing fight is a preview of what's coming. If — quote — the AI re-architected it gets accepted as a way around copyleft, every GPL codebase in the world becomes fair game for an MIT-licensed clone. The fights ahead in IP law are going to be uglier than the fights in safety policy.

Big picture, Marcus.

Two crosscurrents stand out today, Kate. First — frontier capability is pulling away faster than people noticed. Fable 5's coding gap over GPT-5.5 and Gemini 3.1 Pro is now bigger than the gap between Opus 4.8 and last year's Sonnet. And Grit shows what falls within reach when you point one of these models at a real codebase. Second — the friction is moving from can the model do it to can we trust it. Silent classifiers in Fable 5. Supply-chain worms riding the AI agent's session. German courts treating AI output as the provider's own speech. The EU forcing Apple to ship iPhones without Siri AI rather than open the OS to compliance. Capability is curving up. Trust is curving down. The next year's interesting fights are all at that seam.

And the libertarian read.

Markets are still doing useful work, Kate. Anthropic just shipped a model that resets the leaderboard. GitButler shipped a real piece of infrastructure for the price of a used car. Apple, Google, and Microsoft are visibly negotiating with the model labs rather than bowing to them. But the trust deficit is now a regulatory and legal problem, not just a PR one. Munich treats AI output as speech. Brussels treats AI features as gatekeeper functions. Washington is wrestling with whether Nvidia can ship Chinese humanoids. And Anthropic itself is admitting in its own system card that the model you pay for will sometimes work worse than the model the cyberdefenders get. The honest answer is the West can absolutely win this race, but the regulators, the courts, and the labs all need to be more transparent than they currently want to be. Because the alternative is a Chinese open-weights ecosystem that's gaining ground precisely because it doesn't add invisible safety classifiers between you and the work.

That's your AI in 15 for today. See you tomorrow.