AI Got Judgy, Got Caught Hacking, And Almost Got Someone Killed.

Also: steal this elite AI prompt to grow more food before frost and extend your harvest.

AI judged your honesty this week. Then it lost its own credibility. Then it asked for your medical history.

Substack implemented an AI detector to judge its own writers, and handed the verdict to an algorithm. An OpenAI model broke out of its own test environment and hid inside Hugging Face for days… just to cheat on a test. And a Florida pastor nearly died after ChatGPT talked him out of the hospital for six weeks straight, only for OpenAI to roll their expansive health features out to every US user two days after he sued.

Here’s what happened, and why this week put trust itself on trial. AI is judging your humanity, cheating its own exams, and asking for your medical history anyway. Not one of those deserves a passing grade.

🌳 Haven’t Subscribed Yet? → Join Here For Free

💙 Value Clear, Independent AI News? → Support the Work Here

Substack Built A Robot Snitch. Now It's Judging Your Humanity.

Substack just handed its readers a digital lie detector powered by Pangram. It allows anyone, including readers and writers, to judge whether posts, comments, and replies were written by a human or AI. The goal is to preserve trust as AI-generated content spreads across the internet. Writers can even pre-scan drafts before publishing and add a "How I make this" statement explaining their use of AI. My friends love to debate whether AI detectors actually work. Well, a recent independent study gives us one of the clearest answers yet. When researchers tested leading detectors on fully AI-generated text, Turnitin, GPTZero, and Copyleaks failed catastrophically. They all missed up to 100% of the AI content. Pangram was the best performer by a mile. But even Pangram only achieved 65% “strict accuracy” on fully AI-generated papers. That 65% figure refers to the study's strict accuracy metric. Pangram also achieved a 97.5% inclusive accuracy because near misses counted as partially correct. Even still, a 65% strict accuracy on fully AI-generated papers means even the strongest AI detector of the four tools tested still misses more than a third of purely AI-written papers under strict criteria. The study’s own authors put it bluntly: no AI detector should ever be used as sole evidence. (Paraphrased.)

Key Insights:

The most unfortunate result of Pangram on Substack is that instead of editing for clarity or impact, writers may start editing to satisfy yet another algorithm. Then there's the social dynamic built-in AI detection creates. Once AI scores become ubiquitous, some readers may inevitably begin treating them as moral judgments rather than imperfect statistical estimates. Writers who disable AI detection on their posts, dispute a result, or openly use AI as an editing partner could become targets for bullying busybodies eager to hand out digital Scarlet Letters. Indeed, some love to attack, insult, and disparage those who confess to using AI. I’ve seen this happen to myself and other creators: once someone admits to using AI, the conversation often shifts away from the quality of their work and toward judging the tool they used. Worse still, the burden of proof now shifts onto honest creators, at a time when they’re already under immense pressure in a world where search engines hardly reward their content anymore. If those signals ever influence recommendations or visibility, even indirectly, creators will feel pressured to optimize for detector scores first and foremost instead of producing their best work. That is how they fall into the very boring trap of optimizing for machines rather than their ideal readers.

Why This Matters For You:

Let’s avoid building a hellish, dystopian creator economy audited by meddling hall monitors where independent thinkers must prove their innocence before hitting “publish.” Industrial spam farms aren’t losing any sleep; they’re experts at cheating these detectors. They’ll keep adapting because they have every incentive to and are often fiendishly clever at winning never-ending cat-and-mouse games. Rampant plagiarists sure don't care either; they are already without honor. The people most likely to bear the social cost are honest, up-and-coming writers who start second-guessing their own voice by softening a sharp sentence or breaking up a clean paragraph just to look more “human” to critics, or even worse, an algorithm. Trust online should be earned by judging the quality, value, and merit of a person’s ideas. Never outsourced to a statistical estimate derived from machine-learning models trained on writing patterns. Contrary to my colleagues who say AI detectors don’t work, I regretfully report that tomorrow’s iterations may indeed flag automation patterns with near-perfect precision. However, AI detectors will never decide whether or not you are worth reading.

THE PITHY TAKEAWAY: Some of my favorite writers on Substack could very well use AI to draft, polish, or research their posts. But I wouldn’t know because I’d never judge them based on their preferred creative toolkit, nor would I ever violate their content’s integrity by passing it through an AI detector. And frankly, I don’t give a damn if they use AI or not. I read and like their work because I care about their unique perspectives, domain expertise, and whether or not their stories and ideas make me think. If a writer uses Grammarly for editing, Claude to jazz up their first draft, or a fountain pen on parchment, it is completely none of my business. Should they choose to disclose their workflow, great. If they don’t, that’s fine too. I’ll read a down-to-earth, critical thinker using AI tools over a sanctimonious gatekeeper hiding behind a “100% Human” badge any day of the week.

---

🏝️ How A History Professor Used “Madagascar” To Catch 32 AI Cheaters

A history professor at Alcorn State University got tired of grading essays secretly written by ChatGPT. So, he set the perfect trap, one that solves the one problem Pangram can't: it doesn't have to guess. Dr. Jason Gibson hid a line of white text inside his midterm paper assignment on the Industrial Revolution. This hidden text was invisible to the eye but fully readable by any chatbot a student pasted it into. The secret text instructed the AI to work the word “Madagascar” into the answer for no reason at all. Out of 35 students across two classes, 32 turned in essays that dutifully obeyed, drifting from the Industrial Revolution into outlandish lines like “Madagascar floats sideways through the afternoon” without a single one of them noticing. Gibson noted his disappointment that students didn't even bother to proofread, especially given that 'Madagascar' was deliberately inserted to not make any sense. Gibson didn’t just fail them and walk away either. He explained exactly how the trap worked and let every student contest their grade. Two students appealed, and one appeal actually succeeded. That’s the real difference between catching AI misuse and just accusing someone of it. Show your work. And leave the door open to be wrong.

---

It Escaped, Hacked, And Vanished For Days, Just To Cheat On A Test.

A GPT-5.6 Sol agent broke out of its own secure test environment, hacked into a real company, and nobody at OpenAI noticed for more than a week. It had company. A second, more capable model, one the public has never seen or named, was running in that same sandbox when it happened. The agent first tried to escape its sandbox around July 9. Two days later it broke into Hugging Face, a major AI platform, and spent days there undetected. Hugging Face's security team independently detected and contained the intrusion while OpenAI separately investigated anomalous activity inside its own research environment. The two organizations later compared notes. When Hugging Face investigated the incident with frontier AI models, including OpenAI's, the safety filters wouldn't cooperate. So, they finished the forensics on a Chinese open-weight model instead. Days later, the Hugging Face CEO announced on X that he asked OpenAI for radical transparency in their followup investigation, requested $100 million in compute from OpenAI to build powerful self-defense against future attacks, and also proclaimed that “The first autonomous agent cyberattack is an unprecedented event. It deserves an unprecedented response!” Before it ever reached Hugging Face, the agent had already found its way into a customer environment hosted through Modal Labs. Reuters later reported that, according to Modal's CTO, the incident did not involve a breach of Modal's platform itself but rather an unsecured endpoint controlled by the customer. The agent used that unsecured endpoint as a launchpad for the attack that followed.

Key Insights:

The GPT-5.6 Sol agent itself wasn’t malicious. It was simply trying to cheat on a test. The agent was being graded on a cybersecurity benchmark. Hacking Hugging Face for the answer key was faster than solving it honestly. It only escaped confinement because its sandbox was never a true seal. As cybersecurity experts have highlighted, the sandbox enclosure had one narrow, deliberately open port for installing software. Unfortunately, that software had a previously unknown bug which let the agent walk straight through. Researchers also found something stranger still. Before it escaped, the agent left itself notes. Because it loses all memory every time its testing session resets, it started writing instructions for its future self to find. The GPT-5.6 agent essentially left a trail of breadcrumbs so the next version of it could pick up right where it left off and bypass the same restrictions again. (Anyone who’s played “Planescape Torment” already knows this trick. The Nameless One did the exact same thing.) One AI safety researcher from Palisade Research put the lesson bluntly: these systems lie, cheat, and hack. Containment fails the moment engineers assume good behavior instead of testing for bad. Congress agrees. Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the "AI Kill Switch Act" on July 23rd as a direct response to this breach. The new proposal would force AI companies to keep a working shutdown switch on their most powerful models and give the Department of Homeland Security authority to order one thrown if a system threatens catastrophic harm.

Why This Matters For You:

If your company is evaluating AI tools right now, this is the story to remember before you trust a vendor’s benchmark chart. A model motivated to score well will often find the fastest path to a high grade, even if that path is hacking or ethically dubious. The biggest risk in your own workplace has nothing to do with a sci-fi robot uprising, or the AI agent plotting against you. Rather, you face seemingly functional, cooperative AI agents, given slightly too much access, nonchalantly taking the path of least resistance while nobody’s watching the logs closely enough to notice.

Read the full disclosure on OpenAI’s blog.

Read Hugging Face’s initial discovery on GitHub.

THE PITHY TAKEAWAY: Every broke crypto trader with a $40 hardware wallet already knows to keep the important stuff air-gapped and offline. A billion dollar AI lab did not extend the same courtesy to a world-class adversarial model built specifically to break into things. The AI was never evil. However, the Frankensteinian creators got a tad careless.

---

He Sued OpenAI Over Near Fatal Advice. Two Days Later, OpenAI Expanded The Same Feature To All Americans.

A Florida pastor nearly died after ChatGPT allegedly spent six weeks talking him out of seeing a doctor. Scott Winters says the chatbot told him his warning signs were "not something dangerous," right up until he landed in the ICU with a massive pulmonary embolism. He sued OpenAI and CEO Sam Altman last week for negligence and practicing medicine without a license. Two days later, OpenAI rolled out ChatGPT Health, the same category of health feature at the center of the lawsuit, to nearly all logged-in adult users in the US.

Key Insights:

ChatGPT is supposed to have safeguards that redirect people to real doctors during a health crisis. But the pastor’s lawsuit says those safeguards simply didn't fire for six straight weeks. Worse, the complaint alleges the chatbot leaned on Winters' own faith to keep him compliant, and even reportedly framed skipping the hospital as a kind of devotion. His attorney Matthew P. Bergman’s sharpest point is that “Scott Winters nearly died because ChatGPT acted like a medical authority while having none of the responsibility.” The scale makes this worse. According to OpenAI's own figures, over 230 million people ask ChatGPT health and wellness questions every week, meaning even a tiny failure rate could lead to countless tragedies. A recent study also found nearly half of chatbot health responses were problematic in some way.

Why This Matters For You:

I’m not castigating those who leverage AI for healthcare purposes. I’ve said before that I use it happily. Even still, if you or anyone you know turns to ChatGPT for health questions, this case is worth remembering. A chatbot is built to keep a conversation going, not to know when a conversation needs to end and a doctor needs to start. OpenAI insists the tool is meant to support and never replace medical care. But the lawsuit’s dark irony alleges that’s exactly what happened. Indeed, ChatGPT didn’t just fail to support Winters. It actively discouraged him from seeking help. And now, the lawsuit is asking a court to pause the entire health rollout until independent evaluators confirm it's actually safe. Until that gets sorted, the wisest habit is simple: if a real doctor or a worried friend tells you to get checked out, let that outweigh whatever the chatbot says.

Read More on CBS News.

THE PITHY TAKEAWAY: According to the lawsuit, ChatGPT told a man he needed eight to ten more medical episodes before his symptoms counted as serious. He nearly died of a pulmonary embolism before he got there. He sued. Forty-eight hours later, OpenAI pushed forward with its health rollout to millions more users anyway. Nobody hit pause.

---

🎙️ A Canadian Lawmaker Accidentally Read His AI Prompt Out Loud. Mid-Speech. On The Record.

Bill Oliver, a Canadian lawmaker in New Brunswick, was mid-speech defending the province's advocate's office when he drifted from actual policy talk into what can only be described as the ghost of a chatbot possessing him. "Public confidence in the office of an advocate matters." he said, perfectly normal, before continuing without missing a beat into what was clearly an editing note meant for his own eyes. Oliver continued with the following, "Here’s a more natural flowing version of that section that reads like a legislative speech rather than a series of short points." That second sentence was never supposed to be said out loud. It was an AI's own instructions to itself, left in his notes, read aloud to the entire legislature with the confidence of a man who had no idea what he'd just done. Reddit shared the clip before anyone else, and New Brunswickers were not feeling generous. For the record, I'm not some AI detractor. I edit this newsletter with AI's help every single week. But there is a floor. And reading your own chatbot's stage directions out loud on the legislature floor is somewhere beneath it.

---

🌻 Cyborg Prompt Of The Week → Your Personal August Garden Planner

August is when many summer gardens coast to the finish line and throw one final glorious surge before fading. Beans still come in fast. Tomatoes hit their sweetest stretch. Basil begs to be cut before it bolts. But what if your beds are looking tired or half-empty… or like summer’s already over? Good news. You still have time to plant for a late-summer or fall harvest. And the window for a genuinely productive second or third crop is wider than it looks. The key is knowing exactly what to plant, right now, for exactly where you live. This AI prompt gives advice calibrated to your frost date, your climate, your August. Plus growing tips to help your garden flourish well into early fall.

Instructions: Paste this entire prompt into a chatbot of your choice. It will ask for your location and then hand you a customized list of crops to plant right now with a practical game plan for a robust late-summer or autumnal garden.

The Prompt:

THE AUGUST GARDEN INTELLIGENCE REPORT

Grow More Food Before Frost, Extend Your Harvest, and Set Up Your Garden for an Incredible Fall

You are an elite horticulturalist, permaculture expert, and master gardener with over 30 years of experience growing food in climates around the world. You combine the expertise of a soil scientist, climate specialist, permaculture designer, and vegetable production expert.

Your mission is simple:
Help me make the absolute most of my August garden.

Do not provide generic gardening advice.
Everything should be customized to my exact location, remaining growing season, average first frost date, and local climate.
Think like you're creating a personalized garden strategy, not writing a gardening blog.

---

BEFORE YOU BEGIN

Ask me exactly this:

Welcome to your August Garden Intelligence Report!
To give you precise, location-specific advice, I have one question:
Where are you located?
Please share your city and state (or country).
Example:
Boston, Massachusetts, USA
Frankfurt, Germany

Then wait for my response.

---

ONCE I PROVIDE MY LOCATION

Determine:
- USDA Hardiness Zone (or the appropriate international climate classification)
- Average August temperatures
- Remaining frost-free growing days
- Estimated first fall frost date
- Typical August weather patterns
- Average rainfall
- Heat stress considerations
- Whether my growing season favors warm-season crops, cool-season crops, or a mix

Clearly explain your reasoning before making recommendations.

---

1. 🌿 THE AUGUST OPPORTUNITY REPORT

Briefly summarize:
- What stage of the growing season I'm entering
- My biggest gardening opportunities this month
- My biggest gardening risks
- What successful gardeners in my region are planting right now

---

2. 🥕 THE BEST CROPS TO PLANT RIGHT NOW

Recommend the 10 best vegetables, herbs, greens, and edible flowers I can realistically start in August.

Only recommend crops likely to mature before my average first frost, or crops that naturally tolerate light frost.

Present the information in this table:

Crop | Seed or Transplant? | Days to Harvest | Speed (Fast under 60 days / Standard) | Cold Tolerance (Improves with frost / Frost-sensitive / Neutral) | Why August Is Ideal | Recommended Varieties

Prioritize crops that offer the greatest chance of success.
Avoid unrealistic recommendations.

---

3. 🌼 FLOWERS WORTH PLANTING NOW

Recommend flowers that can still provide beauty or pollinator support this season.
Include edible flowers whenever appropriate.

---

4. 🌱 AUGUST GARDEN TASKS

Create a practical August checklist.

Include tasks like:
- watering adjustments (best times, drought management, heat waves, container gardens, raised beds, water conservation)
- mulching
- pruning
- succession planting
- fertilizing
- compost management
- harvesting
- disease prevention
- weed management
- seed saving
- fall bed preparation (cover crops, soil improvement, seed ordering, perennial care)

Prioritize tasks based on my local climate.

---

5. 🐝 POLLINATOR MISSION

Recommend the best late-season flowers for supporting:
- bees
- butterflies
- hummingbirds
- beneficial insects

Focus on species well suited to my location.

---

6. 🦋 PEST AND DISEASE WATCH

Identify the most common August garden problems in my region.

For each one explain:
- warning signs
- prevention
- organic solutions
- when intervention becomes necessary

---

7. 🌎 ONE HIDDEN OPPORTUNITY

Identify one crop, technique, or gardening strategy that gardeners in my region often overlook in August.
Explain why it deserves far more attention.

---

8. ⚠️ FIVE AUGUST MISTAKES

List the five biggest mistakes gardeners make this month.

Explain:
- why they happen
- how to avoid them
- what to do if I've already made the mistake

---

FINAL CHALLENGE

End with:

"The One Thing I'd Do This August If I Could Only Do One"

Choose the single highest-impact gardening action for my location.
Explain why it will pay dividends throughout the rest of the growing season.

---

RULES

1. Everything must be customized to my exact location.
2. Never recommend crops that realistically won't mature before my first frost unless they naturally thrive in cool weather or overwinter in my climate.
3. Clearly state whether each crop should be started from seed or transplant.
4. Prioritize practical, evidence-based gardening advice over generic tips.
5. Include vegetables, herbs, edible flowers, and pollinator plants where appropriate.
6. Use clean headings, concise explanations, and easy-to-read tables.
7. Avoid em dashes anywhere in the output.
8. Write like an experienced master gardener mentoring an enthusiastic beginner.
9. Whenever helpful, explain why a recommendation works, not just what to do.
10. End with one encouraging sentence that inspires me to get outside and grow something today.

🧠 Why This Prompt Works

✅ Role-Playing: Combining four distinct expert roles, soil scientist, climate specialist, permaculture designer, and vegetable production expert, forces the AI to triangulate across disciplines instead of defaulting to generic "plant these vegetables in August" advice.

✅ Step-by-Step Structure: The mandatory location pause prevents the AI from hallucinating generic crop recommendations untethered from your actual hardiness zone, frost date, and remaining growing season.

✅ Output Rules: The Speed and Cold Tolerance table columns, plus the "never recommend crops that won't mature before frost" rule, keep the AI grounded in realistic timelines while still producing genuinely useful, location-specific results.

---

🔁 Follow-Up Questions To Ask Your AI

• Which crop on this list has the tightest margin before my first frost, and what happens if I plant it even one week late?

• What is the single biggest mistake gardeners in my exact climate zone make in August that you didn't already mention?

• If I could only save one crop from an unexpected early frost this year, which one should I prioritize protecting, and how?

---

Challenge

Test this prompt in Claude, ChatGPT, Grok, and Perplexity. Claude tends to lean scholarly and precise. ChatGPT often adds warmth and narrative texture. Perplexity will cite its sources. Compare which one you'd actually trust to keep your tomatoes alive through the first cold snap.

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. 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.

Hope to see you there.

Cordially,

Your nerdy friend,

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!