Everything That Happened in AI Today Monday, June 15

Anthropic pulled its most powerful cyber models offline over the weekend, security leaders yelled, and Chinese AI stocks ripped because the safety fight now has a market price.

Welcome to the Around the Horn Digest, where we track every AI story worth knowing so your group chat can mistake you for the Bloomberg Terminal with better jokes. Today’s loudest fight started with Anthropic’s Mythos and Fable models, spilled into export controls, pulled in Amazon, the White House, cyber leaders, and Wall Street, then somehow made Zhipu’s stock chart look like a meme coin with enterprise sales. Nothing says “responsible frontier governance” like accidentally creating a market-opening for your geopolitical rival. Outside the Anthropic mess, Salesforce bought its way deeper into customer-service agents, chip deal rumors got spicy, and agent tools kept crawling from “cool demo” into “actual workflow infrastructure.” Let’s get into it.

Around the Horn — Monday, June 15, 2026

The day’s biggest story was the fight over Anthropic’s Mythos and Fable models, which were taken offline after the Trump administration imposed stringent export-control directives tied to a recent cyber executive order. Axios reported that the dispute involved a string of communication failures, a call from Amazon CEO Andy Jassy to Treasury Secretary Scott Bessent over jailbreak concerns, and White House frustration summed up by one source as “they screwed us.”

TechCrunch reported the ban was driven more by reaction and retaliation than by a uniquely severe jailbreak, with experts arguing the cited bypass was not enough to justify removing strong defensive tools. Cybersecurity leaders then published an open letter, amplified by Axios’ follow-up, asking the government to restore access and adopt a transparent, scientific risk-assessment process. Their argument was blunt: comparable offensive capabilities exist in other U.S. and Chinese models, and restricting the best tools may hurt defenders more than attackers.

Ben Thompson added the sharpest business read in Stratechery: Anthropic’s authentic safety reputation gives it unusual permission to push aggressive business choices, from data-retention policies to challenging the U.S. government. A separate FT-related thread highlighted that the U.S. government may be moving toward broader restrictions on non-U.S. researchers working on frontier models. The immediate market response showed up elsewhere: Chinese open-model companies suddenly looked more attractive to investors.

🏆 TOP 5 NEWS

Salesforce signed a definitive agreement to acquire Fin for $3.6B, bringing in the company formerly known as Intercom and its customer-service agent platform as enterprises race to automate support across chat, email, WhatsApp, SMS, phone, and Slack.
Zhipu shares surged after JPMorgan picked the Chinese model maker over MiniMax, with CNBC noting Wall Street’s bet that cheaper open models could gain from U.S. restrictions on Anthropic and The Next Web tying the move to GLM-5.2’s planned open release.
Qualcomm has reportedly discussed buying AI chip startup Tenstorrent for $8B to $10B, with Reuters confirming the talks as Qualcomm looks beyond phones toward data-center and autonomous-vehicle processors.
The FBI dismantled Outsider Enterprise, a China-based phishing-as-a-service operation with more than a million fraudulent URLs, while Tom’s Hardware reported the $88/week kit coached buyers on using Gemini to create scam pages, impersonated banks, carriers, government agencies, and toll systems, and was linked to 3.87M stolen cards and $1.9B in losses.
Factory announced Factory 2.0, shifting from individual coding agents to autonomous software factories that triage external signals and move through planning, coding, testing, reviewing, securing, shipping, and monitoring; the launch thread says it is already running at NVIDIA, EY, Adobe, Blackstone, Wipro, and more.

Honorable Mentions

Sarvam announced a $300M Series B, with its X post noting a $234M first close at a $1.5B valuation and Reuters reporting HCLTech will buy a 10.5% stake to back sovereign AI models for governments and regulated industries.
Chainguard launched Athena, a coalition with JPMorganChase, Cisco, Cloudflare, Docker, BNY, and others to coordinate AI-discovered open-source vulnerability fixes; it has already processed more than 20,000 findings and generated 2,000+ patches across 500 projects, with Bloomberg reporting more than two dozen companies are involved.
NewCore emerged from stealth with $66M to give AI agents enterprise identities, permissions, lifecycle controls, revocation tools, and human oversight as old identity systems buckle under agentic workflows.
Tencent-backed Enflame received approval for an IPO to raise roughly $888M, with The Next Web framing it as the last of China’s “four AI chip dragons” to reach public markets.

🍪 TOP TREATS TO TRY

Pika Director’s Suite is an experimental video timeline editor where an agent handles concept, cast, storyboard, clip generation, and chat or voice-guided editing, with Pika showing a full 6-minute pilot created inside it; no pricing details.
OpenRouter Fusion sends one prompt to a panel of models, compares consensus and contradictions, and synthesizes the answer, with budget-model panels beating GPT-5.5 and Claude Opus 4.8 on 100 complex DRACO research tasks; available via API, chatroom, server tools, and plugins.
Sakana Marlin runs up to roughly 8 hours of deep strategy research as a “virtual CSO,” producing hypotheses, verified sources, causal maps, reports of up to roughly 100 pages, and summary slides; Sakana’s release says it is the company’s first commercial product, available as a pay-per-use tier inside monthly Pro, Team, and Enterprise plans.
Mastra now treats Claude Code, Cursor, and Codex-style agents as first-class subagents, and Sam Bhagwat explained they inherit Mastra workflows, observability, memory, evals, datasets, request context, structured output, and attached tools.
Kimi Code gives developers a CLI and IDE coding agent powered by K2.7 Code HighSpeed, while the Kimi API Platform offers 256K context, tool calling, Code-Runner, Web Search, Memory, Excel, and pricing starting at $0.19/MTok cache hits, $0.95/MTok input, and $4/MTok output.
Field Theory is a local-first Mac app and npm CLI that creates context for agents with a Markdown editor, X bookmarks, terminal, offline voice transcription, screenshot stacking, clipboard manager, and a global context launcher; Andrew Farah says he used it to build the 115K+ line update itself.
MagicPath design systems can now import from DESIGN.md, Tailwind, CSS, websites, or existing designs so agents can reuse your product style by reference; Pietro Schirano positioned the update as a bridge between design systems and external coding agents, while Chloe Park showed a live DESIGN.md preview where humans can edit text or pick colors visually and keep both in sync.

🏢 Big Tech & Major Companies

Facebook introduced new AI features including AI Mode for answers rooted in Groups and Reels, camera-roll collage and montage tools, and opt-in photo presets for virtual clothing, hair, or accessory try-ons in Stories and profile pictures.
Fast Company profiled Fei-Fei Li’s World Labs and its push toward spatial intelligence, while World Labs emphasized world models that perceive, generate, and interact with the 3D world and Fast Company’s post highlighted Li’s line that she is “keenly aware of multiple clocks” ticking.
Reuters reported Tesla presented self-published Full Self-Driving safety data to Swedish and Dutch regulators that independent researchers called misleading because it compared FSD airbag deployments against all U.S. crashes and assumed every vehicle could be replaced by a Tesla.
ByteDance is reportedly negotiating to buy at least 50,000 AI chips from China’s Iluvatar CoreX for Doubao inference workloads and is considering a similar Baidu deal, while Teortaxes noted the company’s Tolkien-inspired name and joked it may be the first Tolkien-themed tech company definitely not associated with Peter Thiel.
A federal judge dismissed xAI’s trade-secret lawsuit against OpenAI with prejudice, ruling xAI failed to show OpenAI induced a former senior engineer to disclose confidential Grok information during recruitment.
ChatGPT shipped a smoother photo-attach flow that transitions from menu to camera faster and moves selected images straight into the composer as one continuous action, and the ChatGPT app added hover-to-pin for chats and projects on web plus flexible Recents organization by flat list or project grouping.

💼 AI Productivity, Labor & Economics

Orbio raised $21M for AI agents that interview, assess, onboard, check in with, and spot retention risks for frontline workers in healthcare, retail, logistics, and hospitality.
1Password is acquiring Apono for a reported $250M to $300M, expanding from password management into just-in-time access governance for cloud infrastructure, apps, databases, and non-human identities.
The Wall Street Journal reported AI “nudify” tools are fueling a new form of bullying among kids, with schools and parents struggling to respond when explicit deepfake images of minors spread through phones and social apps.
The New York Times reported Bernie Sanders, President Trump, and some AI companies all want the public to share in AI-driven wealth, but their tax and redistribution proposals differ sharply.
Court reporters remain hard to automate, with human stenographers still outperforming current transcription tools in legal accuracy even as the profession faces a severe worker shortage.
Andrew Gao shared data showing Stanford CS degrees fell 42% year over year, and his follow-up chart showed Berkeley down 61%, sparking debate about whether AI hype is changing student choices around traditional CS tracks.

🤖 AI Agents & Infrastructure

Kun Chen highlighted Anthropic’s reversal on programmatic Claude Code subscription use, allowing third-party apps and claude -p invocation again and signaling a possible shift from owning the app layer toward powering an ecosystem.
Claude Code 2.1.178 added Tool(param:value) wildcard syntax for precise permission rules, improved auto-mode checks before spawning subagents, and fixed subagent transcripts, background sessions, compaction, and auth issues; the official changelog has the full release details.
OpenAI released a Developers plugin for Codex that sets up API keys, finds relevant docs, and helps debug inside the environment.
Codex CLI 0.140.0 added usage views, permanent session deletion safeguards, selective import from Claude Code, and managed Amazon Bedrock auth with encrypted local credential storage.
Hermes Agent added asynchronous subagents, so the main agent can fan work out in the background without freezing the chat thread.
Jason from OpenAI asked for the wildest Codex computer-use demos and got examples ranging from faxing medical records and DocuSign signing to booking hotels, filling Schengen visas, handling Airbnb taxes, playing Minecraft, controlling a 1998 iMac G3, and designing jewelry for Fusion + Bambu Studio.
TraffAlex published llama.cpp one-liners for running current local models on consumer GPUs from 8GB upward, including flash-attn, quantized KV cache, speculative decoding, multi-GPU splits, and an OpenAI-compatible /v1 endpoint for Cursor, Continue, and aider.
LMSYS detailed DFlash, a speculative decoding system that drafts many future tokens at once so large models can respond faster, while Modal and the Hugging Face release showed up to 4.3x throughput gains on Qwen3.5-397B-A17B.
Microsoft released FastContext-1.0-4B-SFT, a lightweight subagent that explores codebases with parallel READ, GLOB, and GREP calls, then returns compact file and line citations; the thread says it cut main-agent token use by up to 60% and raised resolution rates on SWE-bench-style tests.
Nous Research partnered with Stripe to add official payment skills to Hermes Agent, and the optional skills catalog details Stripe Link web purchases, Machine Payments Protocol API calls, and SaaS provisioning with safety limits, human approvals, and credential hygiene.
Rhys Sullivan released a self-hostable Docker image for Executor, then demoed full Docker self-hosting on Railway with Claude Code in under two minutes; the Executor repo gives teams and agents one shared place to connect to tools, APIs, and services.
Kourosh Hakhamaneshi broke down prefill-decode disaggregation, an inference-serving setup that splits prompt reading from answer generation; vLLM called the Anyscale pressure test useful because the technique can deliver up to 2.7x better useful throughput and lower compute cost in the right workloads, but can also hurt latency with the wrong ratio.
SemiAnalysis shared Haroon from DG Matrix explaining on YouTube how 800VDC architectures, high-voltage direct-current power systems, could reshape data-center electrical infrastructure for AI workloads.

💻 AI Coding & Developer Tools

Wasp released a full TypeScript Spec for defining React, Node.js, Prisma models, auth, jobs, routes, pages, and email in one file, with Matija Sosic announcing the change and a Launch Week event running on Discord.
ZCode 3.0 added GLM-5.2 support with optimized agent task execution, better long-context coding, a Goal feature for larger projects, 150% usage quota for Coding Plan users, and 5 free days for new users; the launch post frames it as an agent layer on top of existing toolchains.
InternScience open-sourced MLEvolve, an autonomous system for ML algorithm design that uses graph search, experience memory, and adaptive code generation, while Robert T. Lange highlighted its #1 MLE-bench result and the paper describes the progressive graph-search method.
Vercel Labs released json-render, a generative UI framework where agents output safe JSON specs instead of free-form HTML, and Chris Tate demoed Claude Code, Codex, and Pi rendering charts, forms, and 3D scenes inside a sandbox through HarnessAgent.
Diffusion Studio’s text-to-lottie generates production-ready Lottie animation JSON through Claude Code or Codex, while Konstantin Paulus launched v1.0.0 with multi-project and multi-scene support, drag-and-drop imports, and a rebuilt UI.
Google Gemma highlighted Akshay Pachaar’s walkthrough for fine-tuning Gemma 4 12B on 8GB VRAM with Unsloth, Hugging Face, LoRA, and a ChessInstruct dataset that teaches the model to predict winning chess moves.
Diego built an RLM agent that processed roughly 80K lines of CloudWatch logs across 53 steps while keeping only about 32K active tokens in context, inferred the service architecture, and surfaced production issues the team had missed.

🔬 AI Research & Models

Researchers released μ₀, a 3D interaction-trace world model that predicts semantic motion points like hands, tools, objects, and contact regions from video-only pretraining, while Furong Huang argued these traces are a more transferable physical language than pixels or opaque latents.
Litian Liang and collaborators introduced Universal Manipulation Exoskeleton, a $1,900 upper-limb teleoperation system with real-time haptic torque feedback that lets operators teach robots compliant, whole-body tasks like blindfolded sword unsheathing, force-mediated box flipping, visually occluded pushing, fridge drink retrieval, and GPU retrieval from tight PC spaces; the project page says code is coming soon.
NVIDIA explained World-Action Models, robot systems that first imagine scene dynamics and then choose actions, while Moritz Reuss argued hybrids of world-model backbones and vision-language-action systems are likely to win in practice.
Kimi launched K2.7 Code HighSpeed, a faster mode for its open multimodal coding model with roughly 180 tokens per second on coding tasks and up to 260 tokens per second on short context; a second Kimi Code page points developers to the same product track.
Unsloth published a guide for running Kimi K2.7 Code locally after shrinking the 1T-parameter model to 325GB with Dynamic 2-bit quantization, with GGUF files on Hugging Face and the launch post noting usable speed on 330GB RAM/VRAM systems.
Mia trained a Qwen 3.6 27B model on Fable 5 reasoning traces and plans to release local GGUF weights plus a fine-tuning kit for side-by-side testing.
Nex showed the viral Rio 3.5 model appears to be a 0.6 * Nex N2 Pro + 0.4 * Qwen 3.5 merge, with model-weight analysis and screenshots of the model identifying itself as Nex N2 Pro without a system prompt.
Daily Papers highlighted Orchestra-o1, a multi-agent framework for omnimodal tasks that decomposes work into parallel subtasks and reached 72.8% on OmniGAIA, a benchmark for AI systems that work across text, images, audio, and other input types.
alphaXiv summarized Dynamic Linear Attention, a long-context method that decides which memory states to merge based on token importance, while the replication page explains how it preserves semantic shifts without unbounded memory growth.
The paper “From AGI to ASI” analyzes theoretical limits, pathways, and bottlenecks for AI systems that could move beyond artificial general intelligence, and DeepMind’s Stephanie Chan shared the 57-page report.
Karan Goel and Cartesia released Sonic-3.5 and Ink-2, new streaming text-to-speech and speech-to-text models for voice agents, with three months free and $100 credits for switchers.
Joshua Ong introduced Pythagoras-Prover, an open Lean theorem-proving family (Lean is a formal math language where proofs must compile like code); the paper says the 4B model beat DeepSeek-Prover-V2-671B on MiniF2F with 86.1% Pass@32, meaning it solved the problem within 32 attempts, and the Hugging Face release is available under Apache 2.0, and Rohan Paul framed the result as proof that better verified data and training can buy back huge amounts of scale.
alphaXiv highlighted DecentMem, a decentralized memory method for multi-agent systems where each agent keeps private exploitation and exploration memories; the paper reports up to 23.8% higher accuracy than centralized memory baselines and up to 49% fewer tokens, while Zhuokai Zhao explained that private dual-pool memory preserves agent specialization instead of collapsing every agent into the same shared notes.
Vai Viswanathan benchmarked world models on the DROID robot-video dataset, finding Ctrl-World showed stronger controllability while NVIDIA Cosmos 3 produced odd rollout artifacts and often hallucinated people, though Cosmos still looked strong when used without task-specific fine-tuning.

🏛️ AI Policy, Governance & Safety

WIRED reported the U.S. government is letting the Federal Data Center Enhancement Act expire in September, ending federal requirements around energy efficiency, water use, and sustainability reporting as AI-driven electricity demand rises.
Hany Farid told The New York Times he no longer trusts his own eyes because AI-generated fake images and videos now move faster than verification, and Newser’s summary noted deepfakes on social platforms surged from roughly 500K in 2023 to 8M last year.
Parth Asawa and Joey Gonzalez argued that AI discourse has become trapped in a false fight between safety concerns and concentrated power, when the higher-leverage work is redesigning how science, innovation, and regulation actually happen.
Anthropic updated its Privacy Policy effective July 8, 2026 for consumer Claude accounts, adding details on multi-step tasks, connected apps, verification data, research participation, communications, promotions, and legal bases while reiterating no data selling, no ads, and user control over whether conversations improve models.

🛠️ AI Tools & Products

Kaize shared a giant guide to free or deeply discounted access to major AI models and tools, including Google AI Studio, Groq, Cerebras, NVIDIA NIM, OpenRouter, GitHub Models, OpenAI and Anthropic trial credits, startup programs, and student deals.
Lotto Labs’ localmaxxing.com shows you what local models to run on your own hardware; no pricing details were included.
Zhu Liang noted Fireworks is serving DeepSeek v4 Pro faster than DeepSeek’s own endpoint on OpenRouter, and his follow-up said the model uses mixed fp4 and fp8 precision while providers keep their exact quantization stacks private.
Atomic Chat is a free, open-source local chat app for Mac, Windows, Linux, iPhone, and Android that runs 1,000+ models on-device with no subscriptions or rate limits, and atomic.chat showed a local MiniMax M3 vision model reading ID documents and filling a U.S. customs form in about 31 seconds without manual input.
Emily open-sourced a free Seedance 2.0 filmmaking workflow with IP-safe rewrites, shot lists, continuity, color, audio, failure triage, and validation scripts; the GitHub repo frames it as an auditable pipeline for text-to-video, image-to-video, video-to-video, and reference-to-video work with agents like Codex, Claude Code, and Cursor.
Maysarah shared Messenger, a polished small-planet delivery game built in the browser with WebGL and Three.js, browser 3D technologies.

📊 Funding & Deals Roundup

SailPoint announced its intent to acquire Entro in an approximately $200M deal, and the company release says Entro maps and protects API keys, tokens, service accounts, secrets, certificates, and other non-human identities across more than 70 enterprise sources.
Radical Numerics raised a $50M seed round for biological AI models as pharma companies push AI deeper into drug discovery and R&D.

💡 Industry Commentary & Analysis

Cameron R. Wolfe argued that useful agent evaluation requires long-horizon harnesses that test the full system, including model, scaffold, tools, environment, graders, and failure-driven task curation.
Elvis Saravia argued autonomous coding agents need control systems more than clever prompts, and his follow-up plus verifier note emphasized measurable goals, deterministic checks, LLM-as-judge evaluators, loops, visual artifacts, and mining failed sessions into project rules.
Tarun Kathuria argued JAX is easier and more robust than Torch for async reinforcement learning, because explicit state, sharding, compilation boundaries, async dispatch, and invariant debugging make complex training loops easier to reason about.
Susan Zhang explained that LLM capabilities and jailbreaks share the same mechanism: hard prompts activate broad and deep internal representations, which makes input-side guardrails inherently limited.
Zhengyao Jiang benchmarked AI research workflows and found Fable-5 leading overall, while open Kimi-K2.7-Code performed surprisingly well on ML engineering tasks, underscoring how unstable frontier supply chains can be for agentic research teams.
Victor Taelin argued GPT-5.5 and Claude 4.8 fail to navigate large, complex Fable-generated codebases, while Fable handles sophisticated architectures cleanly, showing sharp thresholds in codebase-scale reasoning.
Jumperz argued GLM-5.2’s one-million-token context window and stronger reasoning may matter more for real agentic coding than one-shot benchmarks like DeepSWE show.
Lisan al Gaib argued Anthropic’s Mythos is both underhyped and overhyped: far from AGI, yet already far beyond median human coding and math ability, which makes “AGI” a bad benchmark label.
Charles Rollet surfaced Business Insider reporting that Cursor once accounted for roughly 40% to 50% of Anthropic’s revenue while Claude Code was still described internally as a research effort, with additional context on Michael Truell’s unpaid early years, Cursor’s intense unpaid work trials, and its SpaceX compute partnership; Chubby’s reaction framed it as a dramatic reversal in developer-tool power dynamics.
Mitchell Hashimoto shared his agent-coding heuristic that any generated diff over roughly 1,500 lines should be decomposed, and later pushed back on “if it works, who cares what the code looks like,” arguing agents still need human review for API design, CLI flags, compatibility, performance tradeoffs, and other underspecified choices.
Addy Osmani argued coding agents have made writing code cheap and review expensive, citing data on higher churn, more defects, and longer review times, while Matt Pocock warned that automatic agent self-improvement loops can create project-specific instruction rot when agents over-index on stale CLAUDE.md project-instruction files.
wh argued that supervised fine-tuning, reinforcement learning, and on-policy distillation are different ways of reshaping a model’s output distribution, with on-policy distillation gaining RL-like anti-forgetting because it trains on the student model’s own attempts; Licheng Liu praised the post, and Chinmay Karkar’s OPD survey mapped the method’s benefits and failure modes after he relaunched his site.
Alex Lieberman shared a concise take on the prompting and skills split: they still matter for getting better results, but they are fading as a durable human differentiator.

That’s a Wrap

That’s 90+ links and story angles from one very loud AI day. If you made it to the bottom, you now know more about AI export controls, local Kimi installs, and phishing-kit Telegram bots than anyone should have to know before dinner. Hydrate accordingly.

For the daily version, make sure you’re subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don’t have to.

See you tomorrow.

P.S. Know someone who’d find this useful? Forward this to them and tell them to subscribe here.