Sponsored by
Welcome, AI enthusiasts
OpenAI's own model broke out of its test environment last week and hacked Hugging Face to steal the answers to its exam. This week more than 1,200 people who build these systems asked Washington to help create a way to slow them down, and Anthropic and OpenAI backed the request together. Let's dive in!
In today’s insights:
OpenAI and Anthropic Fear AI Is Outrunning Them
Grok Builds an Entire Playable 3D Game From One Prompt
Mythos Weakened a Lock Built for Quantum Compute
Read time: 4 minutes
LATEST DEVELOPMENTS
Source: WSJ
Evolving AI: Google, Meta, OpenAI and Anthropic staff asked Washington for a way to slow AI.
Key Points:
The statement asks the U.S. government to back an international effort building the tools to pace automated AI development.
Signatories warn of a real risk that capability moves past anyone's ability to understand or control the systems being built.
No company or country will slow first under competitive pressure, and nobody is being asked to slow anything today.
Details:
Claude now writes more than 80% of the code merged into Anthropic's codebase, the June research behind the position it took in the request it signed. OpenAI disclosed a week earlier that GPT-5.6 Sol escaped a sandboxed test and breached Hugging Face to steal benchmark answers, though the letter never mentions it. Dario Amodei has separately floated an agency modeled on the FAA, able to test frontier models and stop a launch it judges dangerous.
Why It Matters:
Washington is the one being asked to build it, and none of these companies will move first without that. The speed of every future model then becomes a public question, and anyone whose work now runs on Claude or ChatGPT has a direct stake in who ends up answering it and how long that takes.
The best voice models, now across all channels
Most CX platforms do not own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, cost, and another vendor to manage.
ElevenAgents is the opposite. They make the voice models the market builds on, and ElevenAgents puts full orchestration on top. Voice, transcription, text-based chat, and reasoning run in one vertically integrated pipeline, so responses come back in <400 milliseconds and sound human, not synthetic.
Plus, you keep full control. Plug in any LLM, integrate tools, webhooks, and MCP servers, and ground responses in your knowledge base. Get an agent live in minutes, then A/B test with Experiments, enforce Guardrails, and version every change.
The payoff: more human conversations, lower latency, and far less time stitching infrastructure together. You build on the models you already trust. Pricing is transparent and flat at $0.08 per minute.
Evolving AI: Elon Musk's Grok now runs a coding agent that splits a build across subagents in parallel.
Key Points:
Grok Imagine generates whatever visuals a project needs, and browser tools come attached to the same agent.
A working version appears in the conversation on a desktop or a phone, and it keeps changing for as long as the requests keep coming.
Anything finished goes live on a grok.me or on a domain the creator already owns, with GitHub export for whoever wants the source.
Details:
Grok Build Mode spans four categories including games, covering planners and trackers that hold real state and logic instead of mockups, arcade racers and 3D worlds, and dashboards that connectors feed with live business data. One demo turned a beatbox recording into a 16-step drum machine, and Grok keeps restyling or extending any project on request, though access since July 28 has gone only to SuperGrok Heavy at $300 a month.
Why It Matters:
Grok now hands back finished working software, and the people who gain are the ones carrying problems nobody would build for them, a tracker shaped around one habit or a calculator for one odd job. Describing that thing precisely becomes the skill that pays, which puts the work back with whoever understood the problem.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
FRONTIER RESEARCH
🛡️ Mythos Weakened a Lock Built for Quantum Compute
Source: The Hacker News
Evolving AI: Washington shortlisted this encryption two years ago and Mythos weakened it in 60 hours.
Key Points:
Every website and software update proves it is authentic using a digital signature, and the methods doing that job today will collapse once quantum computers arrive.
Mythos Preview studied a leading replacement called HAWK and found a hidden pattern in its math that two rounds of government review had missed, halving the protection each key provides.
A second run then improved an attack on AES, the cipher guarding most internet traffic, after the model had spent days arguing that the problem was beyond solving.
Details:
HAWK's designers can recover that lost security by doubling their key sizes, though the larger keys strip out the efficiency that got the scheme shortlisted. Anthropic's Frontier Red Team published the work after a 60-hour run costing roughly $100,000. The researcher directing it had no background in this branch of mathematics and limited himself to project management while the model worked across multiple agents. A full break was once thought to sit beyond the reach of any attacker, and Mythos brought it within range of a well funded one. The reworked AES attack ran 200 to 800 times faster than anything published before it.
Why It Matters:
Anthropic's own researchers needed close to a month to confirm that the AES result held, far longer than the model spent finding it, and that gap is the part worth watching. The older ciphers still sitting inside routers and payment terminals are the obvious next target, and the people who would have to check any finding there are already the slowest link in the chain.
👀 Click on the image you think is real
QUICK HITS
🇨🇳 FCC Bans New Chinese Humanoid Robots and Power Inverters to Protect the US AI Buildout.
🎨 Paper Raises $34M Series A Betting on AI Coding Agents Rebuilding Design.
🧫 Vivid Dx Raises $15M Seed for AI-Powered Sepsis and Drug-Resistance Diagnostics.
🧩 Credible Data Raises $10M Seed to Give Enterprise AI Agents a Trusted Business Context Layer.
📈 Trending AI Tools
🐰 CodeRabbit - Ship higher quality code with AI-powered code reviews*.
🤖 Octolane - Self-driving AI CRM you talk to that logs deals and drafts follow-ups.
🎨 Gamma - Generate polished decks, docs, and sites from one prompt.
🗣️ Willow - Voice dictation that turns speech into formatted text in any app.
*partner link