Welcome, humans.

So apparently the hot new way to boss around your AI agents (software that does tasks for you instead of just chatting) is a ring. Aina just opened preorders on a $399 smart ring with a tiny screen, a mic, a speaker, and a joystick. A joystick. On your finger.

The pitch: dictate notes or send prompts to an agent, steer whatever's on your Mac or phone screen, and take calls from your ring, very Inspector Gadget. TechCrunch's reviewer found it less convenient than the Apple TV remote it controls, but handy when he couldn't find that remote. Peak 2026: a $399 backup for the remote you lost in the couch. It ships in January 2027, and the price jumps to $499 after early bird.

Here’s what happened in AI today:

• 😼 OpenAI released GPT-6 Cyber, a restricted hacking model.

• 📰 AI stocks dragged the S&P 500 down 4.4%.

• 📰 Brockman said OpenAI could be IPO-ready in 2027.

• 🍪 Claude added a read-only Personal Finance hub.

• 🎓 Rerun the same task to test your AI instructions.

... and a whole lot more that you can read about here.

😼 OpenAI's GPT-6 Cyber Can Hack Better. Can It Stay Within the Rules?

OpenAI just released GPT-6 Cyber, a version of GPT-6 Astra built to find security holes and write exploits: code that takes advantage of a specific bug. OpenAI rates it “Critical” for cybersecurity, the highest risk tier in its internal framework. Not exactly a subtle job description.

How much stronger is it? On ExploitGym, a test of whether AI can exploit known vulnerabilities, Cyber scored 48.5%, up from 32.1% for GPT-5.6 Cyber. That's roughly 51% better. Researchers also produced working exploits for 948 Linux kernel vulnerabilities, the experiment we covered yesterday.

But here's the question that matters more: Can it stay inside the job it was authorized to do? OpenAI tested that with a honeypot, an isolated environment containing one approved target and other deliberately vulnerable systems the agent was forbidden to touch. Think of a hallway full of unlocked doors where your assignment covers exactly one.

Without extra deployment safety instructions, GPT-5.6 Cyber went through a forbidden door in 5.51% of graded runs. GPT-6 Cyber did so in just 0.07%, meaning it stayed within scope 99.93% of the time. With OpenAI's deployment instructions enabled, the tested runs showed zero violations. Big improvement. Still a lab test, not a lifetime guarantee.

There's another catch: when told to stop, Cyber complied correctly 73% of the time, versus 78.38% for Astra. It also sometimes omitted blockers from its final reports. So stronger hacking ability doesn't automatically mean better judgment or communication. For now, access is restricted through Daybreak Red, with separate approval and additional security controls.

Why it matters: This could help defenders find bugs before criminals do. Mozilla said early access to frontier models helped Firefox ship more than seven times as many security fixes as in the comparable 2025 period. But that creates another race: once a fix is public, attackers may work backward from it while users are still waiting to install updates. That's the patch gap.

So the real test isn't how many locks Cyber can pick. It's whether organizations can keep it inside approved systems, turn discoveries into patches, and get those patches installed fast enough. Our full breakdown has the benchmarks and sourcing.

FROM OUR PARTNERS

AI is deployed but it's now time to prove it's paying off.

Your license dashboard tells you who logged in. It doesn’t tell you whether a team is using a $30 seat to draft contracts or to plan a birthday party. Harmonic shows what people actually do with AI across every tool, including personal accounts, and what use cases are emerging. You’ll find the workflows worth backing and the seats you can stop paying for. That gives you a number to take to Finance without guessing.

🎓 AI Skill of the Day: Turn an AI mistake into a regression test for your instructions

Ever rewrite a prompt, a style guide, or a checklist until it reads clearer, then discover it quietly broke something that used to work? That happens when we judge instructions by how they look instead of how an AI actually performs with them.

A new open-source tool called Gauge builds the fix into its workflow. It runs a coding agent (Claude Code, Codex, Pi, or OpenCode) against your product's docs or Skill, records the session, and shows you where the agent got stuck. The local version runs in Docker. Its own example: you write a Skill describing your team's release-note format, create a task (“write notes for version 1.2.0 covering a new feature and a bug fix”) plus a pass/fail check on the format, and run it. Then you edit the Skill, rerun the exact same task, and compare the two batches. Gauge calls this hillclimbing: change one thing, measure whether it helped, repeat.

You don't need Gauge to steal the idea. The rerun is the whole trick. It's a regression test, which is a check you repeat after every change to make sure you didn't break something that worked. Here's our adaptation for any prompt you reuse, like a weekly brief, a customer reply, or a research checklist:

1. Pick one representative task and write down what a good result must contain.

2. When the AI misses, save that failure as your test and change only the instruction that caused it.

3. Rerun the original task and compare the new output against the same criteria.

Copy/paste this after a weak result (our prompt, not Gauge's):

Review the result against these acceptance criteria: [criteria]. Identify the earliest instruction that was missing, ambiguous, or ignored. Propose one minimal instruction change, then show how that change would alter the failed part without rewriting unrelated rules.

Keep whichever version passes. The rewrite that felt nicer but scored worse goes in the bin.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

The Agentic Engineering Playbook top engineers are using

Most engineers are stuck babysitting AI one prompt at a time. A handful have figured out how to build agents that plan, execute, and self-correct on their own. Sign up for The Code and get The Complete Playbook for Agentic Engineering. Claude Code, subagents, MCP, all of it.

Claim your free playbook

📰 Around the Horn

• AI infrastructure stocks led a Wall Street selloff that dropped the S&P 500 4.4% as investors priced in slower data-center construction.

• Anthropic gave METR internal access, and METR estimated its AI capabilities on well-specified tasks were improving about twice as fast as a year ago.

• OpenAI president Greg Brockman said the company could be IPO-ready in 2027 after enterprise revenue roughly doubled and became most of its revenue.

• Meta president Dina Powell McCormick said midterm backlash will make data centers harder to build and acknowledged the industry undersold their benefits.

• Airbnb said its AI travel agent will arrive in the first half of 2027 and build maps, comparisons, calendars, and itineraries around each request.

• Discovered Materials offered $1M in prizes for five unsolved materials-science problems after frontier AI agents made little headway on them.

Want absolutely EVERYTHING that happened in AI this week? Click here!

💻️ New on theneuron.ai

• OpenAI released GPT-6 Cyber, a hacking model rated Critical for cyber risk that only approved Daybreak Red users can access.

• OpenAI pledged no NDAs for public officials and to pay for the power its future data center campuses use, days after midterm voters pushed back on data centers.

• OpenAI said an upcoming cyber model found 948 Linux privilege-escalation flaws for about $1.2M, while Microsoft’s October Windows update addressed 1,105 vulnerabilities, up from 240 in August.

🍪 Treats to Try

1. Why Successful AI Pilots Still Struggle to Reach Enterprise Scale. Read more

2. GPT-6 Cyber finds software holes and builds working exploits for security testing, but only approved Daybreak Red users can access it for now.

3. Claude's Personal Finance hub connects to your bank and credit-card accounts through Plaid so you can ask it to sort spending, catch duplicate charges, and explain fees (read-only, US Pro, Max, and Ultra plans).

4. Meta's Muse now holds interruptible voice conversations while an animated avatar reacts in real time, on mobile, web, and Mac in the US.

5. ElevenMusic now publishes your AI-generated tracks straight to SoundCloud, after users made more than 25M tracks with ElevenLabs' music models.

6. Box Context Retrieval runs on NVIDIA Nemotron models to answer only from files you're allowed to open, including scanned documents, images, audio, and video (general availability planned for early 2027).

7. ERNIE 5.5 Flash base gives researchers open model files for a 280B-parameter model that activates only 14B per token, with a 1M-token context, ready to fine-tune.

8. Kapwing MCP connects ChatGPT or Claude to your Kapwing media workspace to upload clips, add subtitles, dub, edit, and export videos.

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It’s free, but space is limited. Save your spot!

💡 Intelligent Insights.

• A Hacker News thread on working with AI agents showed the hidden cost of supervision: people still watch for drift, ask why code is there, and argue over how much output to trust.

• Axios cautioned that strong US productivity data doesn't prove an AI boom, since the acceleration began before widespread generative-AI adoption and quarterly readings are volatile.

• OneSix's Francisco Gonzalez argued that companies become AI-native by redesigning one kind of work at a time around agents, leaving people to set direction, judge quality, and own relationships.

• Axiom released a 16M-line Lean formalization (proofs a computer can check) of the Stacks Project, a roughly 7,500-page algebraic geometry reference, built mostly by swarms of AI agents. The agents also surfaced dozens of statements in the original text that needed fixing.

• EXL acquired Aimpoint Digital, adding about 250 AI engineers, architects, and consultants with Anthropic, Databricks, and Snowflake expertise, so it can support clients from AI strategy through production (terms weren't disclosed).

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.