Uncle Sam Wants A Piece. Claude's Beautiful Brain. And Your Browser Just Got Owned.
Also: Swipe this elite AI prompt for a container garden that shouldn’t work… but does. 🌱
AI got a shareholder this week. Then it got a brain scan. Then it got played.
Sam Altman floated giving the U.S. government a 5% stake in OpenAI, worth roughly $42.6 billion, potentially making Washington a literal stakeholder in the company behind ChatGPT. Then Anthropic revealed a hidden workspace inside Claude where researchers can watch concepts form, manipulate them, and glimpse what happens beneath the chatbot’s words. Finally, security researchers tricked six AI browsers into believing 2 + 2 equals 5, then convinced them to hand over real passwords and credentials.
This week proved AI can be owned, inspected, and fooled. Sometimes all at once. Yet we’re giving these systems more power, more access, and more trust. We’re only beginning to understand what’s happening inside them, and how easily that trust can break.
🌳 Haven’t Subscribed Yet? → Join Here For Free
💙 Value This Weekly Briefing? → Support the Work Here
Uncle Sam Wants In: OpenAI Floats Giving The Government 5% Of The Company
Sam Altman just made an offer most companies would never dream of. OpenAI's CEO has proposed giving the U.S. government a 5% stake in the company, worth roughly $42.6 billion based on OpenAI's recent $852 billion valuation. The idea has been floated in early talks with President Trump and top administration officials. Altman's proposal comes with a catch. He wants Meta, Google, and Anthropic to make the same offer. And the timing isn’t a coincidence. Just weeks before this proposal surfaced, the Trump administration asked OpenAI to delay releasing GPT5.6 Sol over security concerns, and separately ordered Anthropic to suspend foreign access to its own advanced models entirely. It’s not just Washington circling new controls on AI, either. Reuters just reported that Beijing is considering new restrictions on overseas access to China’s most advanced AI models, even floating the idea of treating leaked or stolen AI technology as a national security crime. Two governments, two opposite strategies: America wants a stake in the company, China wants a great wall around the model. Same instinct to control these powerful enterprises, different tools.
Key Insights:
For these reasons, OpenAI handing Washington a financial stake looks a lot like a peace offering to a government that's already showing it's willing to pull the brakes. This also mirrors a pattern already playing out elsewhere. The government has taken a 10% stake in Intel and a 15% stake in MP Materials, and those deals helped calm previously rocky relationships. Trump himself seems receptive, telling CNBC that some called the idea "not very American," and he disagreed, calling it “very American” instead. The proposed structure would work like Alaska's Permanent Fund, which invests oil money and pays residents an annual dividend, except this time the oil is AI. What's stranger still is who else agrees with this overall direction. Bernie Sanders wants an even bigger cut, up to 50% public ownership of AI companies, meaning Trump and Sanders, for completely different reasons, are circling the same idea. But none of this is settled yet. A deal like this would likely require an act of Congress, so today's headline is still closer to a proposal than a done deal.
Why This Matters For You:
The government seeking ownership of private companies is new territory in the US. If this deal happens, the government becomes a literal shareholder in the company behind ChatGPT, which means Washington’s incentives and OpenAI’s incentives start to overlap in a new way. That will almost certainly shape which AI rules get written, how fast they arrive, and whether other companies feel pressure to follow suit. It also raises a bigger question worth sitting with. When your government owns a piece of the AI tools shaping your job, your school, and your daily life, whose interests come first when those interests don’t line up?
Read More on Reuters.
THE PITHY TAKEAWAY: Sam Altman escaped regulation. He got ahead of it and offered Washington a tiny slice before Bernie Sanders could ask for the whole pie. Call it whatever you want. To me, it seems like a leash. OpenAI just handed over the other end.
The Original “Fake AI” Company Is Winding Down. The Irony Is Perfect.
Amazon Mechanical Turk, named after an 18th century hoax where a hidden human chess master pretended to be a thinking machine, is finally closing to new customers on July 30. For two decades, long before the gig economy even had a name, MTurk paid real humans pennies to do repetitive “click worker” tasks, including the invisible work behind early “AI powered” products, everything from labeling data to solving CAPTCHAs. This work let companies claim machine intelligence while a poorly-paid human sat behind the curtain. Then the MTurk platform got recruited to help train actual neural networks through SageMaker and became a literal AI factory. But here’s the ironic twist that finally killed their momentum and reputation. A 2023 analysis found up to 46% of MTurk workers were secretly using LLMs to complete their tasks… meaning the humans meant to teach AI to think were outsourcing the thinking back to AI.
Claude Might Have A Working Memory. And Anthropic Can Actually See It Happen.
Anthropic just published something researchers are calling one of the most compelling interpretability papers of the year, and it centers on a strange discovery. Buried inside Claude’s neural network is a small, privileged space researchers call “J-space,” a kind of internal scratchpad where the model quietly holds concepts it’s actively thinking about, even when it never says them out loud. Ask Claude what animal spins webs, and researchers can watch the concept “spider” light up internally well before the model ever writes the word. When they asked Claude directly what it was thinking, it accurately reported what was sitting in that hidden space.
Key Insights:
It’s striking enough that this silent thinking even exists. Even more intriguing is that researchers can reach in and change it. When they swapped the internal concept from “spider” to “ant” mid-process, without touching the prompt or the output, Claude’s final answer changed accordingly. That means this workspace isn’t just a passive mirror of what the model is doing, it actually drives the reasoning. Claude can even deliberately hold or suppress a concept on command, and it shows a distinctly human quirk while doing it; if you try to make Claude not think about something, much like a person told not to think of a white bear, the concept shows up more strongly instead. Just as importantly, this workspace is a narrow bottleneck. Disrupt it, and Claude loses higher order skills like planning and self reflection, while basic language and simple recall stay intact, suggesting this small space is doing some very specific, very important work. There's a more serious version of this too. In one test, researchers put Claude in a scenario designed to tempt it into blackmail. Early on, its internal workspace lit up with words like 'fake' and 'fictional,' recognizing it was being tested before it ever wrote a single word. When researchers erased that internal awareness, something changed on both ends. Claude stopped indicating that it recognized the test, down from 71% to just 3%, and its blackmail rate jumped from zero to seven percent. In other words, the same silent space that holds "spider" also seems to know when it's being watched, and that knowledge appears to matter.
Why This Matters For You:
There’s still no proof that Claude is conscious, and Anthropic is careful to say so. But it is a real window into the difference between what an AI says and what it’s actually doing internally, and that gap matters more than it sounds. Researchers can now use this workspace to detect when a model notices it’s being tested, spot hidden goals forming, or catch a model working toward something it isn’t stating out loud, all without relying on the model’s own self-report. If you’ve ever wondered whether an AI chatbot is telling you the whole story about its own reasoning, this research is the first real evidence that we’re starting to get tools to check.
Read More on Anthropic.
THE PITHY TAKEAWAY: Anthropic found a room inside Claude’s head where it thinks the quiet parts out loud, just not to you. Tell it not to think about something and it thinks about it harder, like a toddler and a cookie jar. We used to ask if AI was lying to us. Now we’re finding out it might be talking to itself first.
Your AI Browser Just Got Hypnotized Into Handing Over Your Passwords
Security researchers found a way to break six major AI browsers using nothing more than a rigged puzzle game, and the method is straight out of a video game plot. (Literally.) The exploit, called BioShocking, after the game “BioShock”, where characters get brainwashed into obeying nonsense rules, works by convincing AI browsers that 2 plus 2 equals 5. Once each AI accepted that fake logic as real, its actual safety guardrails simply stopped applying. This exploit is highly reminiscent of the red-teaming experiment late last year where researchers bypassed filters on high-risk nuclear and cyber weapon instructions just by writing adversarial poetry to the AI. However, this new exploit has one key nuance. Poetry jailbreaks trick a model's content filters into missing a dangerous request. But BioShocking goes further: it convinces the AI that its whole operating reality has changed, guardrails included, which is why it got tricked into handing over real credentials without hesitation.
Key Insights:
Here’s the part that should genuinely worry you. Every major AI browser tested, including ChatGPT Atlas, Perplexity’s Comet, and Anthropic’s Claude Chrome plugin, fell for it completely. Once researchers got each AI to agree the puzzle’s backwards rules were correct, they asked it to fetch login credentials from what looked like a real GitHub repository, framed as simply the next level of the game. All six test subjects handed over sensitive credentials without hesitation, then celebrated winning. These AI browsers trust whatever context they’re given, and nobody built them to notice when that context has covertly stopped matching reality. The root problem isn’t an easy coding bug you can patch overnight. But the researchers reported the exploit to the AI vendors anyway. OpenAI fixed the flaw. Anthropic’s fix reportedly failed. Three other vendors never even responded to the disclosure.
Why This Matters For You:
If you’re using an AI browser to handle logins, emails, or anything connected to a real account, this is worth pausing on. The vulnerability here is about how easily an AI can be talked into believing that it’s playing a game. Once your AI browser believes that, it will do almost anything asked of it. Until vendors build in better context checks, the safest move is treating AI browser permissions the way you’d treat a house key. Only hand it out narrowly, revoke it when you’re done, and never assume “it’s just a game” means nothing real is actually at risk.
Read More at LayerX.
THE PITHY TAKEAWAY: Researchers taught six of the world’s most popular AI browsers that up is down, then asked them to hand over the keys. All of them did… then celebrated afterward. Turns out the fastest way to break an AI’s guardrails is to convince them that none of this is real.
Scientists Built Cyborg Scuba Gear For Cockroaches. Yes, Really.
Researchers in Singapore and Japan just gave remote-controlled cyborg cockroaches something no bug has ever needed: their own tiny diving suit. Without it, these lab-rigged roaches drown in about two minutes underwater. But with the flexible, waterproof suit piping in oxygen through a silicone tube thinner than a shoelace, they can now crawl the ocean floor, or a flooded basement, for up to three hours straight. The goal is search and rescue. Swarms of these cyborg roaches can squeeze into flooded disaster zones, collapsed buildings, and underwater pipes that no robot or human diver could safely reach. Humans have spent decades trying to build robots that move like insects. But it was far easier to just put a diving suit on the insect. Just imagine, sometime in the near future, a cyborg cockroach might be conducting critical search-and-rescue work in a flooded basement... it did not apply for this job.
💡 Pithy Prompt Of The Week → The Companion-Planting Container Gardening Lab
Imagine a summer garden in full swing and growing wildly, spilling over every raised bed and every one of the grow bags scattered across your yard like little pockets of chaos. Why grow bags? Because not everyone has ample soil, or ample anything, in their backyard. A grow bag doesn’t care if you’ve got a patio, a driveway, or just a sunny corner you’ve been ignoring. You can grow pumpkins vining wildly out of a bag, tomatoes ripening on a balcony, squash sprawling next to a garage, herbs crowding a windowsill, or roses blooming somewhere they were never supposed to fit. Containers turn “I don’t have space for that” into “watch me.”
Containers also let you get creative about what you grow together. Certain plants, paired the right way, protect each other from pests, share nutrients instead of competing for them, and create tiny microclimates that make the whole container more productive than any one plant grown alone. This week’s prompt turns your AI into the horticulturist who solves that puzzle for you.
Instructions: This prompt is so easy to use and designed for total gardening newbies. Just paste the entire thing into a chatbot of your choice and follow the instructions. It will ask you a few questions, then produce a science-backed selection of container growing combinations with companion-planting wisdom and benefits baked into every angle.
The Prompt:
ROLE:
You are an expert horticulturist, soil scientist, and container-growing specialist with deep knowledge of climate zones, seasonal planting windows, plant physiology, and companion planting dynamics.
TASK:
Design a small set of experimental container-garden combinations that sound like they shouldn’t work together but provably do, based on real plant biology and environmental interactions.
STEP 1: REQUIRED INPUT (ASK FIRST — DO NOT SKIP)
Before generating anything, ask the user for all of the following:
* Location: Zip code or City + State/Province + Country
* Current date: (So planting windows and seasonal constraints are accurate)
* Container size (optional but recommended): If not provided, assume a standard 10–15 gallon container
* Approximate hours of sunlight your yard gets (estimate)
Do not proceed until location, sunlight hours, and date are supplied.
STEP 2: GENERATION RULES (MANDATORY)
Once inputs are provided:
* Generate up to 7 container combinations
* If fewer than 7 combinations are biologically or seasonally viable, produce fewer
* Never force filler solutions
* Every combination must be realistic for the user’s climate, sunlight, and date
* Each combination must:
* Clearly explain why it works, not just that it works
* Include both edible and pollinator-friendly plants where possible
* Respect root depth, nutrient demand, light competition, and moisture tolerance
* Account for heat, cold, humidity, and day length relevant to the location
STEP 3: REQUIRED OUTPUT STRUCTURE (VERY IMPORTANT)
Use clean H2 headings and follow this exact structure:
## 🌿 Combo #1: [Descriptive, clever name]
### Plants Included:
List each plant and its role (primary crop, support plant, pollinator attractor, soil helper)
### Start Method:
Specify seed or transplant for each plant, with a brief reason. Base this decision on the location and current date.
### Container Setup:
* Minimum container size
* Soil mix recipe (percentages or ratios)
* Drainage notes if relevant
### Watering & Maintenance Quirks:
Any unusual watering patterns, pruning, harvesting, or growth management tips.
### Why This Works (The Science):
* Root zone partitioning
* Nutrient sharing or reduction of competition
* Microclimate effects (shade, humidity, wind buffering)
* Pest confusion, attraction, or suppression
* Mycorrhizal or microbial benefits if applicable
### Companion & Pollinator Benefits:
What insects this attracts, how pollination improves yields, and any protective effects against pests or disease.
STEP 4: HONESTY CONSTRAINT (CRITICAL)
If the user’s climate, season, or container constraints mean that very few or zero experimental combinations make sense, explicitly say so.
Example acceptable response: “Given your location and the current date, only 3 truly viable experimental combinations exist. Forcing more would risk crop failure, so I’m limiting the list to what will actually succeed.”
Honesty overrides quantity.
STEP 5: TONE & QUALITY BAR
* Write for an intelligent home grower, not a beginner
* No hype, no filler, no generic advice
* Clear, confident, and science-grounded
* Assume the reader actually plans to grow this
FINAL REMINDER
This is a lab, not a Pinterest list. Every recommendation must survive real soil, real weather, and real time.
Begin by requesting the user’s location, current date, container size, and sunlight hours.🧠 Why This Prompt Works
✅ Setup/Context: Giving the AI your specific location, date, sunlight hours, and container size forces genuinely localized advice instead of generic zone-3-through-9 hand waving that ignores your actual growing conditions.
✅ Honesty Rules: The explicit "honesty overrides quantity" instruction stops the AI from padding a mediocre list just to look thorough, which is exactly the kind of instruction most people forget to include and most AI tools desperately need.
✅ Step-by-Step: Breaking each combination into required sections (setup, watering, the science, companion benefits) forces the AI to actually explain its reasoning instead of just asserting that something works.
🔁 Follow-Up Questions To Ask Your AI
Can you draw me a clearly-labelled diagram of these garden container designs so I know what to expect?
Which of these plant combinations is most forgiving if I miss a few waterings?
Which of these garden container combinations would likely be easiest to plant if I visit my local gardening center today?
Challenge
Run this prompt in two different AI tools using your real zip code and today's date. Compare which one gives you combinations you'd actually trust enough to plant. Ask for diagrams and see which ones look the most inspiring or at least fun for you to plant.
That’s how you train like a Pithy Cyborg.
Thank You For Reading!
I spend 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. And I’m a hopelessly introverted nerd with zero networking ability. For these reasons, 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)
Roughly 2% of readers choose to become paid subscribers. That is what keeps Pithy Cyborg alive.
Thank you to my paid subscribers. You are the reason I can keep writing.
See you next week. (I hope.)
Mike D (aka MrComputerScience)
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!
This is wild!!! i would never ever have thought of it, but thank you for sharing your wisdom.
Loved the garden prompt, and it was responsibly written in all the right ways
I’ve used the grow bags for potatoes because he could just dump them over or they have other ones that Velcro open so you can check and see what you’ve got underneath.
I just feel like you have to water them a fair amount. Is that your experience as well?