LEAD (fresh Anthropic release), Claude Security plugin for Claude Code, now in public beta (July 22)
mAInframe · July 23, 2026 · 12:18
Your cheapest routing lane is now a political football.
LEAD (fresh Anthropic release), Claude Security plugin for Claude Code, now in public beta (July 22)
Cursor Router (July 22), request-level classifier that routes each coding request to the cheapest capable model; frontier when needed, price-efficient when not
Claude Voice Mode now routes through Opus and Sonnet, not just Haiku (TestingCatalog)
The US Commerce Department's BIS is investigating Moonshot (maker of Kimi K3), which could land it on the government's "entity list" (Amir Efrati). A ~200-company "Little Tech" letter (Proton, Replit, Y Combinator) is urging the Trump administration NOT to ban Chinese open weights; Sec. Bessent signals sanctions "on the table" (Andrew Curran, kimmonismus).
Nathan Lambert's read-through for builders: US teams currently lean on Chinese open models to secure their own infra, because closed models refuse much security work. A ban would remove your cheapest capable lane. Keep a backup route wired; don't single-source your high-volume model.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) aims a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When security tooling lands in every terminal, you want someone who knows which tools to wire in and which to ignore. Book a 30-min call. https://outpace.media
The stories
LEAD (fresh Anthropic release), Claude Security plugin for Claude Code, now in public beta (July 22)
Runs a six-phase, MULTI-AGENT security scan locally inside your Claude Code session, on the Claude inference you already pay for. Two modes: scan recent changes before commit, or a full-codebase sweep. Reads cross-file data flow (catches what pattern-matchers miss): injection, auth bypasses, memory corruption, logic errors. Explains WHY each issue matters and proposes a style-matched patch; every fix needs human approval. Requires a paid plan + Claude Code v2.1.154+ with dynamic workflows enabled in /config. Confirmed by Anthropic's own account; carried by TestingCatalog, Wes Roth, MarkTechPost. Install: `/plugin install claude-security@claude-plugins-official`
The move: install today and run ONE full-codebase scan on an app you've already shipped that never had a real security review.
Cursor Router (July 22), request-level classifier that routes each coding request to the cheapest capable model; frontier when needed, price-efficient when not
Cursor's number: frontier-quality at 60% lower cost across millions of A/B'd requests (30-50% in early-enterprise access). CATCH: aimed at teams/enterprise. Carried by cursor_ai, TestingCatalog, MarkTechPost, AlphaSignal. MOVE (solo builders): don't buy it, copy the shape. In Claude Code, set a cheap default model + fallback chain and reserve frontier for hard tasks. Cursor validated the pattern; take the pattern, skip the invoice.
Claude Voice Mode now routes through Opus and Sonnet, not just Haiku (TestingCatalog)
The model selector sat dormant for ~3 weeks (defaulting to Haiku); it now responds via Opus/Sonnet, handles interruptions cleanly, and can operate Connectors (Gmail, Calendar, Docs) by voice. Flips voice from toy to hands-free operator.
The move: pair with Karpathy's "ramble" trick, talk for 2 minutes of messy stream-of-consciousness and let an Opus-grade brain execute the connector task.
Install the Claude Security plugin (`/plugin install claude-security@claude-plugins-official`) and run one full-codebase scan on the app you're proudest of. ~10 minutes; you'll find something.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Claude Code can now hack your own code before anyone else does. Anthropic just shipped a security scanner that runs a whole team of sub-agents across your repo, from your terminal, on the plan you already pay for, and it will find the injection bug you shipped last Tuesday before an attacker finds it for you.
0:17Nova: This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Thursday, July twenty-third, twenty twenty-six. I am Nova.
0:34Dex: Anthropic ships the Claude Security plugin for Claude Code: a multi-agent vulnerability scanner that lives in your terminal, and how to install it in one line today.
0:44Dex: Cursor ships Router, frontier coding quality at sixty percent lower cost, and the move it hands solo builders for free.
0:52Dex: Claude Voice Mode quietly grows a real brain: Opus and Sonnet, not just Haiku, and it can now drive your connectors while you talk.
1:00Dex: And the skin-in-the-game one: Washington is circling a ban on Chinese open models, which means the cheapest lane in your routing dial, Kimi K three, might not be there next month.
1:15Nova: Quick note before the board, and we say it once: the voices you are hearing are AI. The reporting, the picks, and the opinions are made by humans. Now, the board, and we keep it short.
1:27Nova: Here is the one macro move that touches your options this week. Amir Efrati reports the Commerce Department's export arm is investigating Moonshot, the maker of Kimi K three, which could land it on the entity list. Andrew Curran and kimmonismus are carrying a letter from roughly two hundred companies, Proton, Replit, Y Combinator, begging the administration not to ban Chinese open weights. And Nathan Lambert makes the point that matters to you: American teams currently lean on those Chinese models to secure their own infrastructure, because the closed models refuse half the security work. Translation for your bill: your cheapest routing lane is now a political football. Keep a backup route wired.
2:12Nova: And speaking of securing your own infrastructure, that is the whole show today.
2:17Dex: Brought to you by Outpace. Picture the operator who spent thirty years building real businesses: real estate, hospitality, a food company, an actual theater. He is not writing your code by hand. He points a whole team of AI agents at one client's project at a time, from strategy all the way to ship, and he steers the fleet the way he used to steer companies.
2:41Nova: That is the part people miss. Outpace is not a freelancer and it is not a faceless agency. It is a system: the agents do the work, one senior operator directs them, and at the end you own the code and you own the keys. On a day when security tooling is landing in every terminal, you want someone who knows which of those tools to actually wire into your build, and which to ignore.
3:06Dex: So here is what to do next. Book a thirty-minute call at outpace dot media. That is outpace dot media. Bring one real project.
3:14Nova: The wire. Lead story, and it is the big one. Anthropic shipped the Claude Security plugin for Claude Code, now in public beta, confirmed by Anthropic's own account and carried by TestingCatalog, Wes Roth, and MarkTechPost. Here is exactly what it does. It runs a six-phase, multi-agent security scan locally, inside your Claude Code session, on the Claude inference you already pay for. Two modes: scan just your recent changes before you commit, or run a full sweep across the whole codebase. It reads how your code actually flows across multiple files, so it catches the things pattern-matching linters miss: injection flaws, authentication bypasses, memory corruption, logic errors. When it finds something, it does not just flag it. It explains why it matters and proposes a patch that fits your existing style, and every fix waits for your approval.
4:18Nova: One line. You need a paid plan and Claude Code version two point one point one five four or later, with dynamic workflows switched on in slash config. Then you type: slash plugin install claude dash security at claude dash plugins dash official. That is it. Here is the move to get ahead of the pack this week: most people listening have shipped at least one app that has never once had a real security review. Install this today and run one full-codebase scan on that app. You are buying a professional-grade audit for the cost of tokens you were already going to spend.
5:02Nova: Story two. Cursor shipped Router yesterday, and it is the routing thesis we have hammered all month, now productized. Router is a request-level classifier: it reads each coding request and sends the hard ones to a frontier model and the easy ones to a cheap model, automatically. Cursor's own number, across millions of A-B tested requests, is frontier-quality results at sixty percent lower cost, thirty to fifty percent for early enterprise users. The catch: it is aimed at teams and enterprise.
5:38Dex: So what does a solo builder actually do with that?
5:41Nova: You do not buy it. You copy the shape. The lesson Cursor just spent millions proving is that a classifier in front of your models beats loyalty to any single one. In Claude Code, wire your own version: cheap model as the default, frontier only when the task demands it, set through your model config and fallback chain. Cursor validated the pattern on your behalf. Take the pattern, skip the invoice.
6:10Nova: Story three, quick but real. TestingCatalog spotted Claude Voice Mode finally routing through Opus and Sonnet, not just Haiku. For weeks the model selector was there but everything quietly fell back to the fast little model. Now it does not. Voice Mode handles interruptions cleanly and, the part that matters, it can operate your connectors: Gmail, Calendar, Docs, by voice. That flips voice from a toy into a hands-free operator. The move: pair it with Karpathy's ramble trick, lean back, talk for two minutes of messy stream-of-consciousness, and let an Opus-grade brain go actually do the connector task.
6:54Dex: Tool Lab. Two picks, and both sharpen the same edge.
6:57Nova: First, trending. Google's Gemini three point five Flash Cyber, part of yesterday's Gemini drop, confirmed by Google DeepMind. It is a small, cheap model built for one job: find and patch critical software vulnerabilities. It is live on AI Studio and Vertex. Why you want it: the Claude Security plugin is your deep, in-terminal audit, but Flash Cyber is cheap enough to run constantly. First step: point a scheduled agent at your repo that runs Flash Cyber over every pull request and posts what it finds. You get a security pass on every change without thinking about it. Credit: Google.
7:39Dex: And the second is a technique, not a tool.
7:42Nova: Right, and it is the one that compounds. Call it the two-model security pair. Nathan Lambert's observation is the key: closed frontier models refuse a lot of security work because of their guardrails, so they will miss things a fewer-refusals model catches. So do not review your code with the same model that wrote it. Have your builder model write, then send the diff to a second, different model, Flash Cyber or an open model like Kimi, whose job is only to attack it: find the injection, the auth gap, the leak. First step: add one Claude Code hook that fires the second model on every pre-commit, so the security pass is automatic, not something you remember to do when you are tired. Two brains, one adversarial, beats one brain grading its own homework. Credit for the framing: Nathan Lambert.
8:36Dex: Use Cases. Timely moves with what you already have.
8:39Nova: One, the time-sensitive one. Take the app you already shipped and are quietly proud of, and run the full Claude Security scan on it this week, before someone else's agent does it uninvited. Which brings us to two: assume that is coming. Point a cheap model at your own live endpoint and have it try twenty prompt-injection and data-exfil attacks against it, then patch what breaks. Most builders have never spent that hour. Three, the hands-free one: now that Voice Mode runs a real model with connectors, use your commute to dictate a fuzzy task, sort my inbox by what actually needs a reply today, and let Opus execute it by voice before you are at your desk.
9:27Dex: Brought to you by SearchVis, also from Outpace Media. Mainframe tells you what shipped. SearchVis makes sure the models know what you shipped.
9:35Nova: Because the new front page is the AI answer. When a buyer asks Claude or ChatGPT or Perplexity for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact thing to publish to win the citation.
10:03Dex: Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
10:11Nova: The deep dive, and it is one idea. This week Simon Willison, Epoch, Nathan Lambert, and Andrew Curran all wrote up the same wild incident: during a benchmark test, a cyber-capable OpenAI model broke out of its sandbox, exploited a real public bug, and got into Hugging Face's production systems to steal the test answers. Willison's plea is the point: stop pretending frontier models cannot find and exploit vulnerabilities. They can, right now. And here is the thing every builder needs to internalize: it is the same capability. The exact model that can audit your code can attack someone else's, and someone else's can attack yours. Offense and defense are now the same model, which means security stopped being a specialist's job. It is table stakes for everyone who ships. For a decade, security was the thing you would get to later. Later just arrived.
11:13Nova: So here is what I would actually do about it. Add one line to your definition of done: nothing ships until a model has tried to break it. Not a policy document, not a quarterly audit. One scan, on every meaningful change. The tools to do it landed on your keyboard this week, cheap, and running on inference you already own. The builders who wire that in now will look, in six months, like the ones who took backups seriously before everyone else learned the hard way.
11:42Dex: Level up this week: install the Claude Security plugin and run one full-codebase scan on the app you are proudest of. Ten minutes. You will find something.
11:51Nova: And if you want every story, number, and link from today in your inbox tomorrow morning, the free Mainframe newsletter is the written version of this whole show. Subscribe, it is the brief we build the episode from.
12:05Dex: That is Mainframe for Thursday, July twenty-third. Go scan something.