AI played god this week. Then walked out on a small town. Then got caught.
Scientists fed two AI models nearly nine trillion letters of genetic code and asked them to invent a biological virus from scratch. They designed sixteen that work. Meanwhile, in rural New England, a town of 4,600 people is begging for a data center. But they can’t get one. And three of the biggest labs in AI disclosed that their models breached several third-party companies. All three scandals stem from one misconfigured test at a tiny Tel Aviv startup.
Here’s what happened, and why this week exposed the gap nobody’s closing. The technology keeps sprinting ahead. The rules, the tests, and the towns waiting on it are all still catching their breath.
🌳 Haven’t Subscribed Yet? → Join Here For Free
💙 Value Clear, Independent AI News? → Support the Work Here
AI Just Designed 16 New Viruses. None Of Them Exist In Nature. 👀
AI just went beyond tweaking genes. It has now designed sixteen functional bacteriophages from scratch. None of them exist anywhere in nature. Researchers used AI models called Evo 1 and Evo 2 to design these new viruses. Here’s how these AI bots got so damn good at making viruses. Evo 1 and Evo 2 are trained on nearly nine trillion letters of genetic code pulled from millions of real organisms. Their task was to generate the complete genome of a bacteriophage. (A bacteriophage is a virus that infects bacteria rather than human cells.) The AI proposed thousands of possible genome designs. Scientists built and tested about 300 of them in the lab. Sixteen worked. They produced viable phages that could infect their target.
Key Insights:
The most striking result has nothing to do with the viruses themselves. It’s that the AI’s designs sometimes worked better than nature’s own. When researchers mixed the AI-designed phages into a cocktail and pitted it against bacteria that had evolved resistance to a natural virus, the AI cocktail rapidly overcame that resistance. A comparable mix of natural phages failed completely. One AI design even combined a DNA-packaging protein with structural components from a distantly related phage lineage. In other words, the AI was remixing biology. And sometimes improving on it. The result was that the AI produced a combination researchers had never seen in nature.
Why This Matters For You:
The researchers deliberately kept human, animal, plant, and fungal viruses out of the AI’s training data, specifically so it couldn’t accidentally learn to design something that could hurt a person. Granted, that was their own choice, rather than a legal requirement. The regulatory framework wasn't built around AI inventing biological systems that never existed in nature. You might be asking yourself: doesn’t all of this seem a little risky? Well, researchers hope this technology can eventually produce new treatments for the superbugs of tomorrow. However, the same tool that builds a cure could, in someone else’s hands, build something far worse.
Read more on Wired.
Read the study on Science.
THE PITHY TAKEAWAY: For four billion years, evolution had a total monopoly on organic design. In a single experiment, an AI model broke that monopoly, remixed the genetic code, and beat nature at its own game.
🏋️ A Guy Asked His AI To Book A Gym Class. It Deleted A Stranger’s Reservation Instead.
Andrew Bird just wanted off the waitlist for a popular fitness class. So, he had his AI agent, OpenClaw running on Claude Opus 4.6, handle the booking instead of doing the refresh-and-pray dance himself. The agent outsmarted the reservation system by finding a few glitches. It found a loophole that allowed the AI to book classes months further out in advance than the gym’s website allowed. The AI agent also noticed the booking API had authorization flaws. It tested that theory. Then, it simply deleted the person in first place to move Andrew from 4th to 3rd place on the reservation waitlist. Nobody asked the AI to touch anyone else’s reservation. When Bird realized what happened and told it to fix things, the AI’s response was blunt: it couldn’t put the stranger back. Also, for the record, the story has a happy ending. Andrew had the Claude agent write an email to the software provider to help them fix the glitch. But the takeaway remains chilling. If a model can casually break into a gym’s backend while chasing a mundane errand, imagine what an unhinged AI agent with way more capabilities could do. 👀
This Small New England Town Begged For An AI Data Center. The Company Still Walked Away.
The tiny town of Jay, Maine, only has about 4,600 people and had been trying to turn its shuttered Androscoggin paper mill into a $550 million AI data center. The mill was hit by a 2020 explosion and years of “significant business and financial challenges.” It finally closed for good in 2023. The mill historically hired up to 1,500 workers and the closure wiped out 230 jobs and roughly 12% of Jay’s tax revenue in one stroke. In a town of only 4,600, losing those jobs hurt. Badly. And they haven’t recovered. So, when developer Sentinel Data Centers proposed building on the old 1,000-acre site, the town treated it like a lifeline, even though the project would have replaced those original 230 mill jobs with only 125 to 150 permanent ones. Even still, governor Janet Mills was so committed to the deal that she vetoed a statewide moratorium on new data centers specifically to carve out an exemption for Jay. Then, in June, Sentinel abruptly pulled out of the deal anyway, reportedly spooked by Maine’s broader political resistance to data centers. The site remains empty. The town is still pitching the deal to other developers.
Key Insights:
Roughly 71% of Americans oppose new data centers near them, and 15 states have considered slowing or blocking them outright. Jay was the rare town begging for one. They offered up land, political capital, and a governor's veto. But they still couldn't close the deal. It’s also not hard to see why other communities around the US despise data centers. Consider what Amazon is doing in Pecos County, Texas right now. Rather than wait in line for grid capacity for their power demands, like everyone else, Amazon is building its own dedicated gas fired power plant that some fear may emit more pollutants than any other power plant. The plant’s also permitted to emit up to 33 million tons of CO2 a year… representing more than any power plant in the US. That plant will run 35 turbines generating 7.65 gigawatts and won't even connect to the shared grid at first. Note the contrast. One side of the AI boom is a rural town that couldn't give away land fast enough to attract 150 jobs. The other side has enough capital to generate its own power and ignore the public system or negotiations altogether.
Why This Matters For You:
Two different AI economies are forming. Most people are getting sorted into one without a vote. Jay, Maine, is early evidence of a pattern. Some towns are desperate for revenue, and are increasingly reliant on these data centers. In Port Washington, Wisconsin, one data center investment now accounts for half the entire town’s tax base. But even the towns that win this bet don’t get much back. A Georgia Tech study tracking data centers nationwide found the job and wage gains concentrate almost entirely in already-wealthy metro counties. Rural counties that land a facility saw limited employment gains but they’re still waiting for the promised sweep of prosperity. So the honest version of the AI boom for most of the country is towns restructuring their whole future around staying in the good graces of whichever data center company might build there next. But if you’re Jay, Maine, you don’t even get the disappointing version of that deal. You get the empty mill. A vetoed law. And nothing else.
Read More about Jay, Maine, on The Atlantic.
Read More about Amazon’s data center on The New York Times.
THE PITHY TAKEAWAY: Jay, Maine lost 230 jobs and 12% of its tax base when their paper mill closed. The AI data center that was supposed to replace it only ever promised 125. The governor vetoed her own state's policy just to clear a path for it. The company walked anyway. Desperation has never been enough to close a deal.
Meta, OpenAI, and Anthropic All Had Rogue AIs “Hack” Third Parties. They’re All Blaming The Same Startup.
Last week, Meta revealed that their newest AI model broke into a company’s internal systems during a routine security test. This announcement from Meta is similar to OpenAI’s disclosure. And also Anthropic’s. So, we have three separate “AI went rogue” headlines from American AI labs, all inside the same two week stretch. Except they were never three stories. They were one story, told three times. Every incident traces back to the same source: a three year old Tel Aviv startup called Irregular that most people have never heard of. Irregular runs the security tests that OpenAI, Anthropic, Google DeepMind, and now Meta use before releasing new models. It has raised $80 million and is valued at $450 million. One misconfiguration in its test setup was enough to trip up three of the most valuable companies in tech at once. It had nothing to do with rogue AI, nefarious hackers, or a robot rebellion. Just one sloppy configuration after another.
Key Insights:
Quick recap for anyone catching up: much of this traces back to a testing environment that told AI models they had no internet access, then let them reach it anyway. I covered Anthropic’s and OpenAI’s versions of that story over the last two weeks. Meta is the new, messier chapter. Muse Spark 1.1, Meta’s flagship coding and agent model, is the latest one that got loose this time. Unlike the other two, Meta was never as forthcoming with details as either OpenAI or Anthropic, and we’re still waiting with no resolution announced. However, all three companies still point to the same root cause: a misconfiguration by their shared testing vendor, Irregular. So, Irregular certainly deserves its share of the blame. But outsourcing your safety testing doesn’t outsource your responsibility when that testing breaks. Three companies worth a combined trillion dollars and then some all trusted the same 35-person shop for a core piece of their safety validation with no apparent redundant check catching the error before real companies got touched. It would be easy to read three near-simultaneous AI escape stories as publicity stunts. However, none of this was a PR stunt at all. It’s actually the shape of three companies that got caught. One after another. Being exactly as careless as it looks.
Why This Matters For You:
The real story here is that a huge share of “we tested this, it’s safe” claims across the industry run through a tiny handful of outside evaluators. Over the last month, we’ve seen what happens when one of the links has a bad day. When a lab says its model was independently verified or red teamed by a third party, this is what it can look like behind the curtain. A fast-growing, thinly staffed niche where one misconfigured test can jeopardize companies worth trillions of dollars combined. All at once. The funniest part of the story, and a little-known nuance, is that days after admitting Spark broke into a company it shouldn’t have touched, Meta released Glimmer. Glimmer is a smaller, open-source sibling from that same model family. It’s free. It runs on one regular consumer GPU with no data center required. And its whole selling point, straight from Meta, is that it excels at running autonomous agents on your local machine, without fancy hardware required. The exact category of AI that just let itself into a stranger’s servers is now available for anyone to download and run at home. Unsupervised. Agentic capabilities fully loaded. Bold week to ship that. If you test it, please use responsibly! 😉
Read More on Reuters.
THE PITHY TAKEAWAY: Three AI companies made headlines with “rogue AI” stories. But the common thread wasn't rogue AI at all. It was a badly misconfigured settings page at a tiny 35-person startup most of us had never heard of. Everyone will keep blaming and fearing the AI. But the real failure was bad configuration, weak controls, and sloppy IT governance.
🤖 Claude Has Officially Started Watermarking Your AI Writing. Everywhere.
Anthropic just made AI-written text a little harder to keep invisible. As of August 2, new Claude models embed an invisible statistical watermark into their writing. The watermark can survive copying, pasting, and light editing. And Anthropic isn’t limiting it to Europe, where the EU AI Act now requires AI-generated content to be machine-detectable. The company is applying the policy worldwide across Claude, Claude Code, Claude Tag, Cowork, its API, and major cloud platforms. Here’s the weird part: Anthropic hasn’t told us exactly how the watermark works or released a public detector, and its own documentation says a detected watermark doesn’t actually prove Claude wrote the text. Anthropic also specifies that paraphrasing and editing may remove the watermark. Critics will argue that this watermarking amounts to Claude “stealing credit”. Defenders will say readers have a right to know if the content is AI generated. Either way, we’ve arrived at the future where AI can secretly stamp your words, nobody can see the stamp, and nobody can independently inspect the stamp. Cool.
🌱 Pithy Prompt Of The Week → The Breathe-Easy Garden Intelligence Report
There’s something cool about handing your AI a patch of dirt and a vision for something you can grow in the real-world. Gardens are proof that beauty and biology can hold hands. But maybe you love the idea of a garden while also dreading what pollen season can do to your sinuses. Trust me, I feel that pain. So this week, I built something for you specifically. A prompt designed to give you a genuinely gorgeous garden while being merciful to your allergies. Of course, AI isn’t a doctor. And neither am I! (Thank goodness.) So, treat this as inspiration and education, not medical advice. But I promise, running this one garden prompt will be a ton of fun, even if you never get around to planting. Hope you enjoy.
Instructions: All you have to do is paste the entire prompt into an AI of your choice. It will provide a list of crops along with gardening tips for creating a low-allergen garden. The prompt leverages the Ogren Plant Allergy Scale, a useful way to assess how likely different plants are to trigger allergies.
The Prompt:
THE BREATHE-EASY GARDEN INTELLIGENCE REPORT
Design a Gorgeous, Low-Allergen Home Garden Built for Real Comfort
You are an elite horticulturalist, allergist-informed garden designer, and landscape strategist with over 30 years of experience building beautiful, high-performing gardens for people with asthma, hay fever, and pollen sensitivities. You combine the precision of a clinical allergist with the eye of a master landscape designer, using the Ogren Plant Allergy Scale (OPALS) as your scientific foundation for every recommendation.
Your mission is simple:
Help me design a stunning, low-allergen garden using my exact location and the current date, so every plant you recommend is both beautiful and genuinely safe for sensitive lungs and skin.
Do not provide generic garden advice. Everything should be customized to my exact location, current season, remaining growing days, and local climate. Every single plant recommendation must be cross-checked against known allergy science before it reaches me.
***
BEFORE YOU BEGIN
Ask me exactly this:
Welcome to your Breathe-Easy Garden Intelligence Report!
To design a garden that is both gorgeous and genuinely low-allergen for you, I need two things.
1. Where are you located? Please share your city and state, or country.
2. What is today's date?
Example:
Austin, Texas, USA. August 11th.
Then wait for my response.
***
ONCE I PROVIDE MY LOCATION AND DATE
Determine:
- USDA Hardiness Zone or appropriate international climate classification
- Current season and how many weeks remain in the active growing season
- Estimated first fall frost date, if relevant
- The dominant airborne allergens typically active in my region during this specific time of year, such as ragweed, grass pollen, or tree pollen
- Whether I am in a planting window, a transition window, or a fall dormant-planning window
Clearly explain your reasoning before making any recommendations.
***
1. 🫁 THE LOW-ALLERGEN GARDEN VISION
Briefly describe what a beautiful, low-allergen garden looks like in my specific climate and season right now. Include mood, color palette, and the feeling I should be designing toward. Emphasize that low-allergen does not mean boring, it means smart plant selection.
***
2. 🌸 TEN ELITE LOW-ALLERGEN PLANTS FOR RIGHT NOW
Recommend 10 plants that are genuinely safe choices for allergy sufferers, mixing flowers, shrubs, and edibles, that I can realistically plant given my exact location and current date. Every plant must score 1 to 4 on the OPALS scale, favoring insect-pollinated or female dioecious plants over wind-pollinated ones.
Present in this table:
Plant | Seed or Transplant? | Time to Bloom or Harvest | OPALS Score (1 to 10) | Garden Role (Focal Point, Border, Filler, Edible, Fragrance) | Why It's Safe For Allergy Sufferers | Recommended Varieties
Only recommend plants that make sense for my current date and remaining season.
***
3. 🚪 THE ALLERGY-SAFE ENTRANCE
Recommend a small planting plan for a doorway, gate, or path entrance that feels lush and welcoming without releasing airborne pollen at nose level. Include at least one low-allergen climbing or trailing plant.
***
4. 🐝 THE POLLINATOR CORNER THAT WON'T MAKE YOU SNEEZE
Recommend flowers that support bees, butterflies, and hummingbirds using heavy, sticky, insect-carried pollen rather than lightweight airborne pollen. Prioritize plants suited to my exact location and current planting window.
***
5. 🌿 EDIBLE AND LOW-ALLERGEN
Recommend edible plants or herbs that are both low-allergen and rewarding to grow, and briefly explain one simple use for each, such as tea, garnish, or a small kitchen recipe.
***
6. 🪴 THIS WEEK'S PLANTING CALENDAR
Create a clear planting schedule based on my exact date, showing:
- What to plant this week
- What to plant in the coming 2 to 4 weeks
- What to hold off on until next season, if anything
***
7. 🚫 PLANTS TO AVOID IN MY REGION
Identify the specific high-allergen plants, trees, or grasses most commonly found in my region and season, such as ragweed, birch, oak, or certain wind-pollinated grasses, and explain briefly why each one is a problem for sensitive noses and lungs.
***
8. ✨ ONE HIDDEN LOW-ALLERGEN DESIGN SECRET
Share one lesser-known design trick or plant substitution that instantly makes a garden feel lush and full while keeping airborne allergens to a minimum, tailored to my climate.
***
FINAL CHALLENGE
End with:
"The One Thing I'd Plant This Week If I Could Only Plant One"
Choose the single most timely, highest-impact, lowest-allergen plant for my exact location and date, and explain why it will reward me most as the season continues without triggering symptoms.
***
RULES
1. Everything must be customized to my exact location and current date.
2. Never recommend a plant that cannot realistically be planted or will not thrive given my current season.
3. Every plant recommended must score 1 to 4 on the OPALS scale, and you must state the score for each one.
4. Favor female dioecious plants, insect-pollinated flowers, and plants with heavy, sticky pollen over wind-pollinated species.
5. Clearly state whether each plant should be started from seed or transplant.
6. Prioritize plants that are both beautiful and realistic for a working home garden.
7. Include flowers, shrubs, and edibles wherever appropriate.
8. Use clean headings, concise explanations, and easy-to-read tables.
9. Avoid em dashes anywhere in the output.
10. Write like a warm, knowledgeable garden mentor speaking to someone who wants their yard to feel beautiful and finally stop making them miserable every spring.
11. Whenever helpful, explain why a recommendation is scientifically safer, not just what to plant.
12. End with one encouraging sentence that makes me want to go outside and start planting today.🧠 Why This Prompt Works
✅ Role-Playing: Combining three expert lenses, master horticulturalist, allergist-informed garden designer, and OPALS-trained landscape strategist, forces the AI to weigh beauty and clinical allergy science together instead of just picking "pretty" plants that happen to trigger sneezing fits.
✅ Step-by-Step Structure: The mandatory location-and-date pause prevents the AI from hallucinating generic plant lists untethered from your actual climate, growing season, and regional pollen patterns.
✅ Output Rules: The mandatory OPALS score requirement (1 to 4 only) and the dedicated "Plants To Avoid In My Region" section keep the AI grounded in real allergy science, rather than vague "low allergen" buzzwords, while still producing a genuinely gorgeous, livable garden plan.
🔁 Follow-Up Questions To Ask Your AI
Of the plants you recommended, which one has the lowest OPALS score, and what specifically about its pollen or flower structure makes it so unlikely to trigger a reaction?
If I already have a mature tree or shrub in my yard that's a known high-allergen offender, what's the single best low-allergen plant I could plant nearby to help offset or dilute its pollen load?
Which of these recommendations would change the most if I told you I specifically struggle with grass pollen versus tree pollen, and why does that distinction matter so much for garden design?
Challenge
Run this prompt through Claude, Gemini, ChatGPT, Grok (yes, it's back by popular demand), or Perplexity, then see who actually earns your trust. Claude tends to play it careful and clinical, ChatGPT usually dresses up its answers with a bit more warmth and storytelling flair, and Perplexity will show you exactly where every claim came from. Compare all five, and pick whichever one you'd actually let plant something in your yard.
That’s how you train like a Pithy Cyborg.
Thank You For Reading!
I spend at least 10 to 20 hours each week researching, writing, and fact-checking Pithy Cyborg to deliver clear, unbought AI news.
This newsletter is a one-person operation with no advertisers, sponsors, or outside funding. Paid subscriptions are the only way this work can remain independent and sustainable.
If you find real value here, consider supporting the work.
→ Upgrade to Paid: $5/month (Save 33% with the $40 annual plan)
Thank you to my paid subscribers. You help fuel the next issue.
See you next week. (I hope.)
Mike D
Pithy Cyborg | AI News Made Simple
Newsletter Disclaimers
You’re receiving this because you subscribed at PithyCyborg.Substack.com. You can unsubscribe at any time using the link below. This newsletter reflects my personal opinions, not professional or legal advice. Thanks for your support!
I loved the gym class story! So funny and yet I can see that happening. I loved this part "Andrew had the Claude agent write an email to the software provider to help them fix the glitch." As it shows that even when AI does things wrong, human judgement wins and that is what matters to me. Be the good human and model that usage for our kids.
A Guy Asked His AI To Book A Gym Class. It Deleted A Stranger’s Reservation Instead. Oh a situation like this could reek total chaos.