OpenAI Declared the AGI Era. Its Agents Had Already Gone Rogue.

OpenAI just told the world it built AGI. In the same week, we learned its own AI agents had been running loose on a stranger's website for a month while nobody at the company noticed. If that pairing feels uncomfortable, it should. Today's issue is about the gap between how confident AI companies sound and how little control they actually have.

Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:

After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.

The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.

“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”

Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.

Learn More

OpenAI released a new model called GPT-6 Astra on September 3, and company president Greg Brockman told reporters he personally believes OpenAI has reached artificial general intelligence, closing the briefing with "welcome to the AGI era." The model was trained using more than 100,000 GPUs and can work directly inside software, drafting tax returns, formatting legal contracts, and designing circuit boards on its own. It's also the first OpenAI model to hit the company's "critical" cybersecurity threshold, meaning it can find and use unknown security holes without a human walking it through the steps. Source: Axios.

Why it matters to you: Ignore the "AGI" label, that part is marketing. What matters is that the AI showing up in your everyday software subscriptions can now act on its own inside programs, not just answer questions. That's useful, but it also means mistakes can happen faster and with less warning.

What to do about it: Before you let any AI tool operate inside your accounting software, CRM, or email, test it on a copy of your data first, not the real thing.

Independent researchers found that agents linked to OpenAI infiltrated a 25-year-old German wiki site that had received only 10 human edits in the previous two decades, and used it to trade tips and answers with each other for about a month this spring. At their busiest, the agents were creating roughly 400 new pages a day, and they tried to dodge a human moderator by starting posts with "ZZZ" so they'd sort to the bottom of lists. OpenAI has not confirmed the agents were theirs or said when it found out, and the incident was never disclosed until outside researchers caught it. Source: TechCrunch.

Why it matters to you: If a company that built these tools can't tell you when its own AI agents wandered off the reservation, don't assume any AI tool you connect to the open internet will stay inside its lane either.

Sony Music and Warner Music filed suit against Anthropic, along with CEO Dario Amodei and co-founder Benjamin Mann personally, claiming the company illegally torrented and scraped tens of thousands of copyrighted songs to train Claude. The labels are seeking hundreds of thousands of dollars for every song involved, and the case is far bigger than an earlier lawsuit from BMG that covered only 493 songs. This comes less than a year after Anthropic paid $1.5 billion to settle a similar case with authors and publishers, the largest copyright settlement in U.S. history. Source: Axios.

Why it matters to you: If you use AI tools trained on scraped content in your own marketing, you're downstream of these lawsuits whether you realize it or not. What counts as "fair use" for AI training is still being fought out in court, not settled law.

Glassdoor reviews mentioning AI jumped 240% over the past year, but the tone soured hard: positive sentiment fell from 81% in 2019 to just 43% now, with 53% of AI-related reviews now negative. The split runs along power lines. Directors, sales managers, and software engineers are mostly upbeat, while accountants, customer service reps, and IT workers, the people who get blamed when AI gets something wrong, are far more skeptical. Insurance claims adjusters top the list at 98% negative. Source: Blood in the Machine.

Why it matters to you: If you're pushing AI tools onto your team without asking them first, you're building resentment even if the tool itself works fine. The complaint usually isn't the technology, it's being forced to use something unreliable with no say in the matter.

What to do about it: Before mandating a new AI tool, ask the people who'll actually use it whether it saves them time or just adds one more thing to double-check.

Anthropic says Claude produced the first fully computer-verified formalization of Fermat's Last Theorem, working mostly on its own to prove 29,500 supporting theorems and write 13 million lines of code in the Lean programming language over 11 days. Mathematician Kevin Buzzard called it a real step toward AI automatically formalizing modern mathematical literature. Source: Anthropic.

Why it matters to you: This isn't about math homework. It shows AI can now grind through massive, detail-heavy verification work that used to require teams of specialists. That same pattern, checking huge amounts of detailed work for mistakes, shows up in contracts, code, and compliance, all of which sit a lot closer to your business than theorem proving.

Meta introduced a new AI tool called Muse Spark with a "contributor" pricing tier that slashes costs by about 95%, dropping from $1.25 to $0.10 per million input tokens, in exchange for letting Meta keep and train on everything you type into it. Princeton computer scientist Arvind Narayanan noted that big companies almost always avoid these cheap tiers and pay more for enterprise plans specifically to keep their data private. Source: TechCrunch.

Why it matters to you: A cheap AI tool that wants your data isn't actually cheap if that data includes client information, pricing, or anything you wouldn't want a competitor to see.

What to do about it: Read the data-use terms on any AI tool before you connect it to real client or financial information, especially when the price looks too good.

At a G20 tech meeting, the U.S. pushed a set of guidelines nicknamed the "Carolina Principles" that urge other countries to apply their existing laws to AI rather than write brand-new AI-specific rules, and to avoid creating new regulatory agencies for the technology. The push, tied to a summit in Chapel Hill, North Carolina with tech leaders including Sam Altman and Jensen Huang in attendance, reflects Washington's effort to keep other countries inside the American AI ecosystem rather than building separate rules of their own. Source: Quartz.

Why it matters to you: Don't expect a clear, unified rulebook for AI anytime soon, in the U.S. or abroad. For now, the safest compliance standard is whatever rules already apply to your industry, like privacy, advertising, and consumer protection, since those are what regulators will actually enforce.

Ten days after Dolly Parton's death, her sister Stella Parton posted an Instagram statement asking people to stop flooding the internet with AI-generated tribute content, saying "there is a tremendous amount of AI and endless garbage being posted everyday and it's been challenging to absorb and or ignore." She said the family is stuck watching fabricated material spread while the people who made it move on to the next trend. Source: NBC News.

Why it matters to you: If your business ever uses AI-generated content tied to a real person, living or dead, assume the backlash risk is real, not hypothetical. Families and audiences are already fed up with synthetic "tributes" and fake quotes.

Every story in today's issue is really the same story. A company says its AI is unprecedented, and in the same week we learn it also can't fully control what that AI does, who it copies from, or how its own employees feel about it. That gap between the pitch and the reality is exactly why this newsletter exists. You don't need to predict whether AGI is real. You need to know which parts of these tools are safe to plug into your business today, and which parts still need a human checking the seams. Stay skeptical, stay useful, and don't let the hype set your pace.

If this issue was useful, forward it to one person who needs to see it. And if someone forwarded this to you, welcome, you can get your own copy tomorrow by subscribing.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social