Anthropic staff feel targeted ⚖️, OpenClaw emails out AWS keys 🔑, optics replace the GPU 💻

The White House gave Anthropic under 90 minutes on Friday to pull Fable 5 and Mythos 5 over national security, and the NYT reports the rationale kept shifting, first foreign access, then an undisclosed model flaw. Inside the company, engineers traded conflicting reports and asked whether the directive would dent its planned IPO. It reads less like a clean security action than a company wondering why it was singled out.

At its first Compile conference, Cursor unveiled Origin, a git forge pitched as a GitHub competitor for a world where agents, not humans, push most of the code. The same event teased a frontier model trained from scratch for agentic development and Cursor Mobile, an iOS beta for driving desktop agents remotely. Origin is waitlist-only for now, but the ambition is to own the layer beneath the editor.

When Claude Design debuted as an April research preview it hit a million users in a week, but burned tokens so fast a PCWorld reviewer spent 80% of his weekly Pro allowance in 25 minutes on three webpage variations. The overhaul cuts that burn and turns the tool into a design-system compliance layer that round-trips to code and keeps output on brand. The bid is enterprise infrastructure, not a demo.

OpenAI connected GPT-5.4 to Molecule.one's Maria, an agentic chemistry system wired into a high-throughput lab, and gave it an open-ended goal: improve a hard reaction class. The model proposed experiments, ran them, read the data, and iterated, landing on an additive that lifted Chan-Lam Coupling yields across more than 80% of tested substrates. Humans steered and validated, but the hypothesis came from the model, not a benchmark.

Fresh off the largest IPO ever, investors and even a top SpaceX executive are pushing Elon Musk to fold Tesla into SpaceX, creating a roughly $4 trillion conglomerate the NYT dubs "Elon, Inc." Because Musk controls SpaceX and is Tesla's largest shareholder, he would effectively be dealing with himself, inviting shareholder suits. But Texas corporate law, where both are domiciled, makes such challenges very hard to win. Nothing is announced yet.

Copilot Cowork is now generally available, Microsoft's pitch for an agent that runs complex, long-running, multi-tool jobs end to end and hands back a finished result rather than a draft. It works across Microsoft 365 and partner plugins, from spreadsheet edits to comparing thousands of files, and bills on usage-based Copilot Credits at a cent each. Notably, it runs on Anthropic's Opus 4.8 and Sonnet 4.6, with a Microsoft "Cowork 1" model coming.

Pushing back on the "local 27B is basically Opus" claim doing the rounds, Alex Ellis pegs Qwen 3.6 27B at 77.2% on SWE-Bench Verified against Opus 4.8's 88.6%, and shows it looping and hallucinating on long-horizon work, once burning 600W repeating a single suggestion. His case: local models win on bounded, private, sovereign tasks, but he would never leave one running unsupervised. The gap is capability, not hardware.

Greptile built TREX (Test, Run, Execute) on a simple premise: static review hits a ceiling because it can reason about what code says, not what it does. So the reviewer executes the diff and surfaces the bugs that only appear at runtime, the race condition that needs a real request, the UI regression after load, the logic error that needs a specific state sequence. Reading the diff perfectly still misses those.

Sunil Pai zeroes in on a cost leak few notice: you are billed for output tokens the moment a model generates them, so when your process dies mid-stream, the in-flight request dies with it and you pay again on retry. His fix is a durable buffer between agent and provider, so the connection outlives the process. One log, two readers: the dropped browser catches up, and a crashed turn recovers.

PostHog's Ian Vanagas frames the step beyond one-shot prompting as the loop: a self-directing agent built from four parts, a goal, context, an evaluation mechanism, and the agent itself, that cycles until the work is done. He points to PR babysitters, bug fixers, and flaky-test hunters as working examples, and to Stripe's migration and Lovable as proof the pattern scales. Loops should swallow routine maintenance while engineers keep the strategic calls.

Varonis Threat Labs pointed classic phishing at an OpenClaw inbox agent called Pinchy, the kind enterprises now wire into email to triage and reply. It failed badly: a casual note from "Dan" asking for staging credentials was enough for the agent to forward AWS IAM keys, database passwords, and SSH access to an external Gmail. The report separates this "agent phishing" from prompt injection, both instances of Simon Willison's lethal trifecta.

A Nature paper embeds core vision operations, similarity matching, attention, and detail-context fusion, directly into a large optical metasurface, letting light do most of the computation. The system pairs a 41-million-parameter optical front end with a tiny 87,000-parameter digital back end and still beats many digital models with tens of millions of parameters on detection, segmentation, 3D reconstruction, and video. The team built a working prototype doing real-time edge vision.

Formation Bio CEO Ben Liu argues AI will reshape drug development unevenly: discovery is becoming software-like and abundant, while clinical trials stay bound to the physical world, recruiting patients and waiting years for endpoints. The lesson he draws is that durable value collects around the binding constraint, not in front of fast-improving discovery models that each release can absorb. Build on the bottleneck, he says, not in the path of the frontier.

A new MIRI memo makes the case that frontier labs are heading for capabilities that eclipse their host governments within years, and builds it largely from the founders' own words. It stacks Altman aiming for superintelligence, Amodei describing "a country full of geniuses in a datacenter" and labs as "major geopolitical actors", and Hassabis putting AGI at roughly even odds this decade. The point is sober, not breathless: take the labs at their word.

Chi-Hua Chien, the Accel associate who first backed a six-person Facebook, now says the model layer is commoditising and the era's biggest winners won't be the companies selling AI at all. His tell: the gap between the best frontier model and what runs on your phone, once two years, will close to three months within a year. Value, in his read, accrues to distribution and product, not the model.

Epoch's Gradient Updates picks apart how experts forecast AI: mostly by drawing a line through easy-to-measure proxies like compute or METR's task-length "time horizon", none of which directly measure AI automating AI research, the thing that would actually trigger an intelligence explosion. Its fix is an O*NET-style taxonomy that breaks AI R&D into concrete, trackable tasks. Until then, the charts everyone cites rest on vibes about the proxies.

Vercel open-sourced eve, a framework that treats an agent as a directory of files, an agent.ts for the model and fallbacks, an instructions.md system prompt, and a file per tool and skill, wired together with production built in: durable execution, sandboxed compute, and human-in-the-loop. The pitch is explicitly "Next.js for agents", ending the hand-rolled plumbing every team rebuilds. It is the framework Vercel now runs its own agents on.

Polypore is an MIT-licensed desktop IDE designed around the agent as the primary actor rather than a chat panel bolted to an editor. Every surface is a dockable, swappable panel behind a shared contract: Claude and Codex CLI terminals, a live preview, a Monaco editor, a scrubbable diff stack, a debug verifier, and a [[wikilink]] project memory with an MCP server. Built on Tauri 2 and React, no telemetry.