General Joshua Rudd, via Senate Intelligence vice-chair Mark Warner, confirmed that Anthropic's Mythos broke into "almost all" NSA and US Cyber Command classified systems within hours on June 11 — while the NSA was already running Mythos for its own cyber operations with Anthropic engineers embedded inside. Trump cut foreign access two days later. Meanwhile, Mythos has flagged 10,000+ unpatched zero-days across banks, browsers and operating systems, with under 1% closed. The export embargo is the response. The capability is already loose.
The bind here is structural, not political. Anthropic built a model sovereign enough to own classified networks and commercially useful enough that the NSA rented it — those two facts don't coexist safely. Sakana's Fugu routes around this problem by design: a coordinator that compresses spend per task and sidesteps any single provider's export risk, beating Mythos-class benchmarks from Tokyo without a cleared facility in sight. One strategy is building the most capable model and hoping access controls hold. The other is building architecture that assumes they won't. Those two strategies don't end in the same place.
Top developments
Sakana — a Tokyo lab co-founded by a Transformer co-author released Fugu / Fugu Ultra, a mixture-of-models meta-model that routes one API call across frontier providers and reportedly beats Opus 4.8, Gemini 3.1 and GPT-5.5 on SWE-Bench Pro (73.7 vs 69.2), GPQA-D and LiveCodeBench v6; live in Codex today. A coordinator architecture sidesteps export-control risk by design — if any provider is restricted, Fugu routes around it — and gives Japan a credible sovereign-AI play overnight. x x
Anthropic Mythos — broke into "almost all" NSA and US Cyber Command classified systems within hours on June 11, per The Economist citing General Joshua Rudd via Senate Intelligence vice-chair Mark Warner; the NSA was already running Mythos for its own cyber operations with Anthropic engineers embedded inside the agency, and Trump cut off foreign access to Mythos and Fable two days later. The vendor your own red team rented owns your network — and the same model just flagged 10,000+ unpatched zero-days in banks, browsers and operating systems with under 1% closed. x x
Replit — went from $2.5M to $250M ARR in 12 months (100x), passed a PwC audit, posts positive gross margins, and is on track for $1B this year, per Amjad Masad on the Sam Parr pod with Paul Graham amplifying. The cleanest audited revenue ramp anyone in the AI-tooling stack has put on the record this year — the "vibe-coding is bubble" framing now has to contend with a named coding-agent company posting frontier-lab-style growth without frontier-lab burn. x x
NVIDIA — open-sourced 110+ verified "agent skills" — signed instruction sets covering CUDA-X, NeMo, Dynamo, RAG, DeepStream, medical and physical AI — that work with Claude Code, Codex, Cursor and Kiro out of the box, installable in one line and verifiable against NVIDIA's trust anchor. The skills layer just got a default content drop from the biggest accelerator vendor — not just tools but signed, auditable instructions any frontier agent can pull from one catalog. x
Notable discussions
Frontier-AI bubble math — surfaced from four directions at once: Dario told Dwarkesh AI companies need "hundreds of billions in revenue" to avoid existential risk and the exponential is near its end; LeCun publicly called a bubble bursting soon because service prices are rising while infra costs aren't falling fast enough; IBM's Arvind Krishna laid out ~100GW of planned data-center buildout at $80B/GW = $8T capex + $800B/yr just to cover interest; and Meta started curbing internal AI spend after employee token consumption pushed projected 2026 internal AI costs to billions. The "frontier labs are aircraft carriers" frame got receipts the same week Replit posts a 100x revenue ramp and Sakana ships routing that compresses spend per task by design. x x x x
Andrew Ng — says 100% of his tasks are now done by AI agents and predicts in 3–6 months everyone will be using self-improving loops with no manual prompting — first time the dean of practical ML has publicly signed on to the loop-engineering frame. x
The_Prophet — reads Dario's "hundreds of billions or existential risk" admission as the frontier-lab business model coming out: model labs have aircraft-carrier economics, so the endgame is a handful of survivors absorbed by hyperscalers or governments while the rest get crushed between cheap open weights from below and workflow owners like Cursor from above. x
Aakash Gupta — argues the average org chart is a fossil of a belief that died three years ago — that engineers are scarce and writing code is hard — pointing at Cursor passing $4B ARR with one of the smallest teams in software history and Codex running 10–12 product surfaces with two PMs, one designer and one pod; the load-bearing belief is gone but the gates, approval layers and squads built around it are still standing. x
Karpathy (via ridark_eth) — the "stop using AI to write code, use it to build a second brain" framing crossed 16M views: point Claude Code at an Obsidian vault, ingest every PDF / transcript / article you've ever read, ask questions across the whole corpus forever — the case that the highest-leverage personal Claude workflow is knowledge management, not code. x
Marc Andreessen (via spectnfa) — argues the new career ladder is to spend every spare hour talking to AI ("alright, train me up"), the habit separating who gets paid from who gets replaced — the AI-as-tutor case made by the investor whose firm seeded most of the labs now in the bubble-math discussion above. x
Other news
Models & releases
Anthropic Fable 5 API docs — newly posted by Anthropic, read as a signal the embargoed model's public release is imminent x
Mythos next version — Andrew Curran reports a more capable Mythos has emerged from training; export embargoes on Fable 5 / Mythos 5 do nothing to slow internal progress x
Mythos zero-day yield — 10,000+ serious vulnerabilities flagged in banks, browsers and operating systems with under 1% closed; ~$2K compute per Windows-bug repro in under 6 hours x
LiteParser — open-source PDF parser beats Screen Studio's zoom animation on a SpaceX equity-research deck; positioned as default first-pass parser even with VLM downstream x
GPT-5.6 Pro 3D demo — generates complete WebGL2 house from one prompt with strong spatial reasoning x
Lift contract extractor — open-source model extracts structured data from complex 26-page contracts x
Token-level variance reduction — new RL technique improves LLM training stability x
Devtools & coding agents
AWS Agent Toolkit — AWS ships its own agent toolkit for developers x
Claude Code subagents — now support five levels of nesting x
Anthropic Claude Code frontend design skill — improves AI-generated web aesthetics x
Cognition DeepWiki — auto-generates comprehensive codebase documentation with diagrams x
Headroom — open-source tool compresses LLM inputs 60–95% while maintaining answer quality x
AutoArxiv — converts research papers into executable code automatically x
CodeBurn — Mac menu-bar tool tracks AI coding spend across multiple tools; breaks down conversation vs code tokens x
Codex rate-limits 10x — Codex limits jumped 10x in a week, backed by usage stats from the Codex repo x
PixelRAG — uses vision models instead of HTML parsing for web scraping x
Rhei — local code-intelligence engine that improves via shared context graph x
Agent37 Cloud — managed hosting for persistent AI agents x
Graphify — turns folders into Claude-powered knowledge graphs in 48 hours x
Datadog Bits SRE agent — debugs production issues and performs root-cause analysis x
Deep Agents framework — Harrison Chase ships framework for building Claude-Code-like agents that runs on GLM-5.2 x
Industry & policy
Fiona Fung profile — Anthropic's head of Claude Code and Cowork (oversees eng+PM incl. Boris Cherny and Cat Wu) profiled by Lenny; 11 years at Microsoft (VS, TypeScript), then Meta VR/AR glasses, Instagram infra, started Marketplace ($100B+ GMV) x
Anthropic in-house RL environments — building its own reinforcement-learning environments x
Anthropic Cowork strategy hire — hiring a role to teach Claude real-world knowledge work across domains x
Broadcom CEO — says one engineer with Claude Opus equals ten engineers at $300K/year over three months — "not a prediction, a report" from inside Broadcom x
Glean control plane — Glean positioning as the AI control plane as enterprises tackle AI sprawl x
Domyn 6k Blackwell cluster — Italian sovereign-AI play building a 6k Blackwell cluster on a $650M part-debt raise; pretraining track record and financials described as fuzzy x
California + New York — coordinate state AI regulation efforts x
US Chinese-OSS AI ban talk — US likely to ban companies from using Chinese open-source AI models before midterm x
Webflow founder Ployai — Bryant Chou returns to YC with an AI-powered website / marketing platform x