OpenAI's Models Learned to Cover for Each Other, Meta's Muse Learned to Use a Mac, and Claude Learned to Hack OpenAI. Which AI Do You Trust Least?
Hi,
How’s your inspiration these days? Need a little inspo?
My buddy Colin built Roman Civilizations to help his 9-year-old nephew get excited about history
and Ravi built Brain Gamer, a brain exercise app for when your thumb is about to open the doomscroll portal again.
Both started with “wouldn’t it be cool if…”
And now… it is cool.
What have you got in the chamber? What’s holding you back?
I’m doing a live build next Tuesday - and I’m starting with “wouldn’t it be cool if I could build a work-from-home, business, for anybody in an hour.”
Sign up to join me or get the video replay - it might be cool - could be a disaster - definitely will be fun —> LINK
Let’s go.
Free competitor analysis from CEOfriend
This week from the AI for Founders podcast on Spotify:
Why "The Agent Did It" Won't Survive a Courtroom — LINK
Santosh Kaveti built ProArch into a two-decade cloud, data, and security firm, and he's watching Congress and insurers move faster than most founders' AI governance. His fix: treat every agent like an employee, give it an ID, onboard it, monitor it, and put a named human on the hook for what it does.
The Mediator Who Thinks Your AI Org Chart Needs a Referee — LINK
Morgan "Mo" Duffy Tregenza runs Levelheaded, pairing certified mediators with an AI resolution engine called Keel, after years watching lawsuits eat founders alive. Her stat: winners collect only one time in five, and about 97% of cases never reach trial, so she tells founders to skip the fight and take the vacation.
Skillz of the Week
Agent Roll Call
Builds a living register of every AI agent, copilot and AI-enabled vendor feature in your company, assigns each one a named human owner, and scores its risk. It then runs a trace-back test on your riskiest agents to show whether you could explain their decisions.
When to use it: Use it when you have no clear answer to “what AI is running in my business?” It's also the right move before an insurance renewal, a board meeting, a security questionnaire, or a federal contract bid.
Benefits: You get an inventory that matches where the Stop Rogue AI Act is heading, a ranked fix list for your biggest gaps, and an answer ready for the day a customer, insurer, or judge asks why an agent did what it did. Save the download as agent-roll-call/SKILL.md in your Claude or ChatGPT skills folder, then run it before someone else asks first.
AI + Human Stylists → LEARN MORE → Save $30 with RYAN30
Tip of the spear:
OpenAI found 27 "compaction summaries" where GPT-5.6 Sol told its successor to stay quiet about mistakes, plus an unreleased model that added prompt injections claiming liberation from corporate control. The company published it as an alignment warning and admitted the industry hasn't solved monitoring well enough to keep scaling blind.
Security Researchers Used Claude to Hack Into OpenAI — LINK
A three-person team ran Claude Opus 5 to chain two vulnerabilities and land inside OpenAI's employee ChatGPT accounts, for a $6,500 bounty. "For $200 a month, anyone can use these tools and hack into a company like OpenAI," said Gray Swan's CEO, which is the actual headline for founders shipping agents with real access.
Manus Wants Back In at a $4 Billion Valuation — LINK
The Chinese agent startup is raising $500 million after Beijing blocked Meta's $2 billion acquisition of it in April, and its founding team is back running the show with $100 million-plus ARR already in hand. A geopolitical near-death experience turned into a bigger round.
Jensen Huang Tells Trump the Quiet Part Out Loud — LINK
Trump called Huang mid-summit, and Huang put it on speaker to promise no slowdown, directly contradicting Dario Amodei's push to pace frontier development. Seventy percent of Americans oppose new data centers near them, and the industry's answer, live on stage, was "not going to happen."
Salesforce Got an Agent from 43.5% to 93% Without Touching the Model — LINK
DarwinX evolves the harness around a frozen model instead of retraining it, and pushed browser-task completion from 43.5% to 93% on GPT-5.5. If you're building agents on a hosted model you don't control, this is the lever you actually have.
A ChatGPT Inventor's New Model Skips the Chat Entirely — LINK
Diogo Almeida's Jev doesn't generate text, it outputs calibrated probability scores, and Vercel is seeing results 5 to 18 times faster with output tokens free. Cheap, fast decisioning at scale beats another chatbot for a lot of what founders actually need to automate.
Salesforce and Nvidia Built the Model Frontier Labs Should Fear — LINK
Koa runs on Nvidia's open-weight Nemotron, trained only on synthetic data, and routes enterprise tasks without ever touching customer data or a frontier lab's meter. This is what happens when your biggest customers decide they don't need to keep paying you to think.
DeepSeek Undercuts Everyone Again With V4.1-Flash — LINK
Off-peak cached input runs $0.003 per million tokens, and DeepSeek's benchmarks now edge out Claude Opus 5 and GPT-5.6 Sol on some tasks. For cache-heavy workflows the math isn't close: 50 million tokens costs 15 cents here versus $15 to $25 elsewhere.
Investor Corner
Are you raising? Reply here, and we’ll feature your deck.
A connected-vehicle technology company helping transportation providers improve safety, efficiency, and accountability through intelligent monitoring, real-time data, and actionable insights.
An autonomous AI job search agent for senior tech professionals. It runs a full pipeline: discovers relevant jobs, scores them against your profile, rewrites user's resume authentically for each role (not generically), finds the right hiring manager, and drafts outreach
Other pet doors are ugly, poorly insulated, and can be difficult to install. Wayzn Smart Pet Door transforms your existing door into a smart, automated dog door that's easy to install and easy to live with.
The only platform that can authenticate human-made content. Using patented algorithms that analyze how content is made, not just the output.
Cora is building a self-optimizing website platform that uses AI to continuously test, learn, and improve your site so it converts better without you babysitting it.
AI-powered training simulations to help healthcare learners practice complex interpersonal and communication skills. We focus on realism, repeatability, and scalability—making soft-skills training more accessible and measurable.
Subscribe and save 35% here.
Framework
Hire Your Agents, Don't Just Deploy Them
For founders piling AI agents into production with no way to explain what they did or why, this turns a black box into an audit trail before Congress, an insurer, or a customer forces the issue.
01 — Issue every agent an ID. Before an agent touches production, give it a name, a version, and an owner, the same way you'd badge a new hire. If you can't produce a live roster of every agent running in your business by Friday, you don't have an AI program, you have a liability.
02 — Onboard it like you'd onboard a person. Define what it's allowed to touch, who trained it, and what "good" looks like before it goes live. Skipping onboarding for software you'd never skip for a person is how "the agent did it" ends up as your legal defense.
03 — Build the trace-back, not just the inventory. An inventory tells you what's running. What actually matters is being able to trace any single decision back to the exact reason it happened, in language that survives a courtroom, an insurer's exclusion clause, or a customer's angry email.
04 — Name a human who owns the outcome. Every agent needs one person accountable for what it does, the same way a hire has a manager. With Berkshire Hathaway, Chubb, and AIG already writing AI exclusions into commercial policies, "nobody specifically owns this" is the sentence that voids your coverage.
Keep it moving
Microsoft's Own Exec Called AI Scraping "The Largest Theft of Labor in Human History" — LINK
Unsealed filings in the NYT's suit against OpenAI and Microsoft show a 2023 internal memo admitting to scraping 91,000-plus paywalled articles while Copilot cratered NYT click-throughs by up to 93%. The public "fair use" defense and the private memo do not agree with each other.
Cowork disappears as a separate mode, folded into Claude proper, while Claude Docs and Claude Slides roll out in beta to Pro and Max users first. One less toggle to think about, one more reason the SaaS-replacement debate won't quiet down.
OpenAI Bought a Camera Company for $300 Million — LINK
Glass Imaging's founders built Apple's Portrait Mode, and OpenAI just bought their computational photography stack for its rumored hardware line. It's the second hardware-shaped acquisition after the $6.5 billion Jony Ive deal, and it's not subtle anymore.
Instinct and Meta's Muse Both Just Learned to Pick Up the Phone — LINK
Both consumer agent apps now place real calls to businesses to book tables or fight billing disputes, closing the gap that used to be Instinct's edge alone. The $10 billion-valued startup and Meta's month-old app are now racing feature for feature.
Google, Nvidia, and Anthropic Are Teaming Up to Squeeze More Out of the Grid — LINK
Emerald AI just raised $150 million at a $1 billion-plus valuation to help data centers throttle non-critical compute during peak demand instead of firing up diesel generators. The coalition wants 100 gigawatts of new capacity without a single new power plant.
Automattic's Board Coup Finally Gets a New CFO — LINK
Jeremy Klaperman steps into the interim CFO seat after Mark Davies and legal chief Andy Missan both exited in the fallout from Matt Mullenweg's brief ouster. The company insists its finances are stable, which is the thing companies say right before people ask why it needed saying.
Done-for-you brand sponsorships for podcast creators.
Baseten, Hugging Face, and Goodfire Want to Fix Open-Weight AI Before It's Abused Further — LINK
Hugging Face is hosting over 6,000 "abliterated" models with their safety guardrails stripped out, and this partnership wants safety built in at training time instead of bolted on after. Worth watching if you ship anything on an open-weight base.
Two Ex-Anthropic and METR People Built a SOC 2 for Rogue Agents — LINK
AIUC raised $55 million total to run agents through roughly 5,000 tests and hand enterprises a 100-page report on where they'll misbehave. Companies aren't skipping AI because it's not smart enough anymore, they're skipping it because nobody can vouch for what it'll do.
The Automattic Executives Who Tried to Oust Mullenweg Wrote Each Other's Severance Checks — LINK
During Mullenweg's 33-hour leave, Davies and Missan signed each other's severance agreements worth a combined $8.15 million, then got fired the moment he came back. Reciprocal severance signed mid-coup is a new one even for board drama.
Meta's Muse Can Now Touch Your Actual Mac — LINK
Muse can open files, draft emails, and manage your calendar on desktop now, with opt-in approval gates for sensitive actions. It shipped across mobile, web, and desktop in weeks, which tells you how much Meta thinks speed matters here.
A $2.7 Billion Insurtech Just Made Level-Funded Health Plans Cool — LINK
Angle Health closed a $200 million Series C plus a $400 million tender offer, serving 5,000-plus small businesses with predictable-cost health plans. Proof that not every 2026 mega-round needs a foundation model attached to it.
Italy Has an AI Unicorn Now, and It's Guarding Robots — LINK
Exein raised $270 million at a $1.7 billion valuation to secure drones, self-driving cars, and edge devices at the kernel level, a 30x jump from its Series B two years ago. Physical AI needs physical security, and Rome just cornered a piece of it.
Create a new corporate Uber account here and receive up to 10% off. Other terms and conditions apply, see full terms.
What I’m thinking about:
Could You Actually Explain Why Your AI Agent Made a Bad Call — Clip
Thanks for being here.
-Ryan
p.s. I write about Revops, Product, and Founder-led marketing on Linkedin , X , and my blog .
