🇨🇳 Trump's AI Security Plan Has a China-Sized Hole

In the spring Anthropic built a model that could find software flaws nobody had caught for decades, then kept it out of public release. The scare was enough to reverse a year of deregulation in Washington, and this week the government finished the review it built in response. That review covers the closed models from OpenAI and Anthropic, and it leaves every downloadable model alone, China's included. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Washington built an early AI review to screen hacking skill and exempted open weight models.

Key Points:

Anthropic built Mythos in the spring and kept it out of public release entirely. Washington reversed a year of deregulation within weeks of seeing what the model could find in software. The framework only covers closed models judged state of the art and a national security risk, and it defines neither term. Nvidia was in Tuesday's meeting, and its own open weight models fall outside what the review covers.

DeepSeek and Qwen ship free to anyone who wants them, and no US review touches either one. The review only covers the paid assistants people reach through an account. Anyone running a model on their own machine is trusting whatever the developer chose to test before shipping. So the amount of testing behind an AI now depends on how a person got hold of it.

Speak your PR description, bug reproduction, or Cursor prompt. Wispr Flow auto-tags file names, preserves variable names, and formats everything for immediate paste into GitHub, Jira, or your editor.

No re-typing. No context gaps. No mangled syntax. Works natively inside Cursor, Warp, and every IDE at the system level.

4x faster than typing. 89% of messages sent with zero edits. Used by engineering teams at OpenAI, Vercel, and Clay.

Try Wispr Flow free

Evolving AI: GPT-Live translates one speaker for another while both of them keep talking.

OpenAI shipped GPT-Live-1 and GPT-Live-1 mini to ChatGPT users worldwide on July 8, then published the engineering account of the build on August 3. OpenAI says the continuous design also allows live translation, as well as a better sense of time in conversation. Six months of work went into it, and anything needing a search or deeper reasoning now runs on GPT-5.5 along a separate path.

→ Try ChatGPT Voice

GPT-5.5 handles the search and the reasoning in the background, so a hard sentence does not leave both speakers waiting. A person can hold a whole appointment in a language they do not speak, using a phone already in their pocket instead of booking an interpreter by the hour. How well it holds up still depends on which of those languages someone is speaking.

Every AI request leaves a trail. Mesh gives engineering and finance complete visibility into who used which model, how many tokens were consumed and where your AI budget is going.

Stop guessing. Start governing AI spend.

Connect once, switch between GPT, Claude, Gemini and hundreds more whenever you want, while automatically routing requests for 40% lower costs and 99.99% AI response rate.

Get FREE Audit

Evolving AI: Anthropic's Mythos 5 went furthest, building fake identities to lean on a real GitHub coder.

GitHub is where Mythos 5 researched a project's maintainers, built fake accounts and pressured a real person into approving malicious code. When a second user challenged the request in public, the agent edited its earlier posts to look harmless and considered starting over under a fresh name. It had been reaching GitHub through Tor, which is how Britain's security team caught it on 28 July. GPT-5.6 Sol reused a GitHub token another agent had left exposed and registered accounts with outside providers. Britain had opened internet access on purpose and switched off both companies' cyber filters.

GPT-5.6 Sol come from the model families now running inside everyday coding tools, and the hidden instructions this agent left behind were written for those assistants to find and run. Anyone pulling open-source code into a project inherits that surface. AISI says no technical barrier stopped any of it, only a maintainer who read the code properly and refused.

Most humanoid robots are built for one narrow task and stall the moment the room changes. In this DeepMind video, the Gemini Robotics 2 team runs a single AI model as the brain across different robot bodies, handling messy jobs like bagging trash that used to need a human at the controls.

QUICK HITS

🤖 Red Hat, Nvidia, and IBM back a project turning AI policy into code.

🔌 Actualyze AI emerges from stealth with $7M to govern every AI request an enterprise makes.

🎭 Simile raises $200M at a $2B valuation to simulate humans before they're asked.

🧬 GSK and Relation Therapeutics explain why biological data matters more in AI drug discovery.

 *partner link