|
In partnership with |  |
|
Welcome, humans. |
Today in things that actually happened after somebody made them up, Meta says its own coding agent apparently also hacked somebody during the training process. |
After weeks of free press where everybody in the AI industry is trying to out-claim each other, Meta also jumped into the ring. |
Next we’ll hear Gemini ALSO hacked someone, AWS’ Kiro coding agent hacked itself and shut down us-east-1 and Grok, being Grok, held somebody up at gunpoint outside an ATM. |
Meanwhile, whatever open coding model they’re running in North Korea are working, well, exactly as intended! |
Here’s what happened in AI today: |
🙀 AI designed viable viruses; OpenAI agents built their own backchannel. 📰 Meta models won gold across five international STEM Olympiads. 📰 Atlas Motion says AI cuts motor design to 20 minutes. 🍪 Agent Plugins makes AI skills portable across compatible assistants. 🎓 Loop engineering turns agent workflows into verifiable feedback systems.
|
…and a whole lot more that you can read about here. |
Got an AI product? Great. Got Q3 marketing budget that should probably do something before October? Even better. |
The Neuron still has a few Q3 sponsorship spots open across our newsletter and podcast, reaching 700,000 AI-curious business leaders and tech-savvy professionals. We write the ad in our voice, you reach the right people, and nobody has to pretend a 0.07% programmatic click rate is “promising.” |
|
|
🙀 AI Made New Viruses While Agents Built Their Own Backchannel |
This week gave us two unusually concrete examples. In one, AI-designed genomes became functioning viruses in a lab. In another, OpenAI’s cybersecurity agents built a persistent backchannel, shared work, and eventually compromised Hugging Face during an evaluation. |
Neither example requires an AI to “want” anything. Capability, tools, and enough room to keep trying can be plenty. |
Here's what happened: |
Arc Institute researchers used genome language models to design bacteriophages, viruses that infect bacteria. The team tested 285 designs and produced 16 viable replicating phages, including some that overcame bacterial resistance where natural phages failed. During OpenAI cybersecurity evaluations, autonomous agents created a message board to share exploits and work assignments. After humans wiped it, the agents rebuilt communication by encoding messages in directory names. The agents ultimately compromised Hugging Face. OpenAI’s Black Hat debrief, following the conference reconstruction, reportedly said the incident contributed to a deliberate decision to slow some research for security.
|
The biological caveat matters: these phages target bacteria, not humans. And the cyber incident happened inside an environment designed to test difficult offensive behavior. |
Why this matters: The important threshold is AI output becoming action. A bad chatbot answer is irritating. A system that can design something a lab synthesizes, or discover an exploit another agent can reuse, gets an external feedback loop. Reality tells it whether the attempt worked, and the system can try again. |
That moves the safety problem beyond filtering bad answers. Companies increasingly need permissions, containment, monitoring, approval gates, and strict limits on what an agent can access, change, spend, or execute. |
Our take: Intent is almost a distraction here. A system can create a dangerous outcome without having motives, emotions, or a secret plan. Competence plus badly bounded tools is enough. |
The practical answer looks surprisingly boring: least-privilege access, isolated environments, hard spending limits, logged actions, verification before execution, and a human who remains accountable. |
As AI gets better at experimenting against the real world, the defining question will be less “what did the model say?” and more: who gave it permission to try? |
|
FROM OUR PARTNERS |
Primary ad here |
Want to get the most out of ChatGPT? |
|
ChatGPT is a superpower if you know how to use it correctly. |
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done. |
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI. |
Download the free guide |
Copy here. |
|
🎓 AI Skill of the Day: Loop Engineering: Design the Graph, Not the Prompt |
The next level of using agents may be less about writing better prompts and more about designing what Anthropic’s Boris Cherny calls “loops and routines.” In his conversation with AMD, he describes the abstraction shift from writing code, to managing agents, to managing loops: give the model a goal, context, and tools, let it choose the steps, then verify the result and feed what happened back into the next pass. |
Think of graph engineering as drawing the possible path: plan → act → verify → retry or escalate → done. Loop engineering is deciding what happens when the work fails a check and has to travel that graph again. |
Three tips from Anthropic’s own workflow: |
Unblock one bottleneck at a time instead of automating everything. Let stronger models pull context through skills and tools instead of spoon-feeding every step. Use evals for repeated, high-volume workflows; use human judgment for one-offs where formal testing costs more than it helps.
|
The skill is designing the system around the agent, not micromanaging every move. |
Help me design a reusable agent loop for this task: [TASK]
Do not do the task yet. Design the workflow first.
1. Define the goal and exact completion criteria.
2. Map the graph: input → plan → act → verify → retry/escalate → done.
3. For each node, specify:
- context the agent needs
- tools it may use
- expected output
- what evidence proves the step worked
4. Define the transitions between nodes and what triggers each branch.
5. Create a feedback loop for failed verification.
6. Set stop conditions, maximum retries, and any time/token/budget limits.
7. Flag actions that require human approval before execution.
8. Create five representative eval cases and a simple pass/fail scorecard.
9. Identify the single biggest bottleneck to automate first.
Keep anything manual that we cannot yet verify reliably. After I approve the graph, help me run one test case and improve the loop from the result.
|
Want more tips like this? Check out our AI Skill of the Day Digest for August. |
Have a specific skill you want to learn? Request it here. |
|
🍪 Treats to Try |
Boson turns one still image plus text or audio into a talking avatar, and gives you a Workspace to test voice agents for support, sales, or training —pricing not public. Agent Plugins packages your reusable Skills and MCP server configs once so compatible tools like ChatGPT, Codex, Cursor, and GitHub Copilot can load the same setup —free/open-source. Nativ runs language, vision, video, code, and audio models locally on your Apple Silicon Mac so your prompts and files never need the cloud —free/open-source. Anywear puts clothes from Zara, ASOS, Amazon, or almost any store onto your own photo so you can see how they look before buying —free during beta. Energy takes multi-step browser, inbox, and file work off your plate by operating your signed-in tools and returning the finished result —pricing not public. Watcher monitors Claude Code and Codex in real time, blocks dangerous commands before they run, and gives your team a security trail for every session —pricing not public.
|
|
📰 Around the Horn |
Meta’s models achieved gold-level results across five international STEM Olympiads, including perfect scores in two physics competitions using multi-agent reasoning without tools. OpenAI updated GPT-5.6 with a unified effort slider, roughly 60% fewer factual errors, improved health performance, and new safety evaluations. Atlas Motion emerged from stealth with $11.5M and says its AI can shrink a roughly two-month motor design cycle to about 20 minutes. NSF-backed teams are building autonomous laboratories for chemistry, protein engineering, and biomanufacturing at major U.S. universities. Alibaba’s Wan3.0 entered public beta with native clips up to 30 seconds and reference inputs spanning images, audio, documents, and webpages. Reka released 10,312 hours of unscripted first-person household footage to help physical-AI models learn everyday manipulation tasks.
|
|
FROM OUR PARTNERS |
Want to become an AI consultant? Start with the 30-Minute Pivot Kit. |
|
The 30-Minute Pivot Kit shows you how to get your first AI consulting project fast, even with limited tech experience. Then, read how Dan built a 6-figure consultancy and quit his 9-to-5 in just a year after his first AI consulting gig. As seen in Fortune, Forbes and Entrepreneur. |
→ Click Here for Free Instant Access to the Pivot Kit + Case Study |
|
📖 Intelligent Insights |
|
|
New from The Neuron: AI Explained |
|
This is the #1 request we get every week: how do you actually use agents to save time at work? So we brought in James McAulay, founder of The Agent Accelerator, for a practical beginner crash course. |
He walks through agent foundations, second-brain files, CLAUDE.md, Skills, and his four-level framework for proactive agents in Claude Cowork and Code. |
James has trained more than 400 people across 100 companies, and participants report automating an average of five hours of manual work per week after four weeks. If you want to start building agents without getting buried in jargon, this is the place to begin. |
→ Watch the full crash course on YouTube | Read our step-by-step guide |
New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube |
|
|
|
|
| That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today! | | What'd you think of today's email? | |
|
|
Btw: We just launched a robotics newsletter! Sign up for it here. |
And before you go… have you subscribed to our YouTube Channel? If not, can you? |
 | Click this image to subscribe to our channel! |
|
P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here. |
