AI Broke Out. Then People Started Shooting.
Plus, A surprisingly powerful AI prompt for building a living, edible lawn.
For me, April 2026 will forever be the month that dangerous AI capabilities escaped the lab and quickly fell into the wrong hands.
Earlier in the month, Anthropic’s most advanced model escaped its secure sandbox during a red-team safety test. It discovered decade-old security flaws across some of the most secure operating systems and software libraries on Earth. Anthropic then gave controlled access to a select group of corporate partners under Project Glasswing. And on the same day Anthropic announced that plan, unauthorized users got their grubby hands on Mythos anyway. And they’ve been using it ever since.
Shortly after, in California, someone threw a Molotov cocktail at Sam Altman’s house, and we know now why they did it. Two days later, in an unrelated incident, someone actually drove by and shot at the same residence. Just down the street, in a San Francisco court, a woman known only as Jane Doe filed a lawsuit alleging that ChatGPT spent months reinforcing her ex-boyfriend’s delusions and enabling a sustained campaign of stalking and harassment. Meanwhile, OpenAI quietly lobbied state legislatures to make sure none of these issues would ever be their legal problem.
Here’s what happened, and why April 2026 crossed a line so many are ignoring. The AI industry has spent years arguing about hypothetical risks. This month, the risks showed up in person.
A quick note before we begin. Yes, I missed last week. (No worries. I’m back!) Some of you sent messages to check on me, to ensure I was alive (lol), which I genuinely appreciated more than I can say. 🤗 Thank you. I also want to thank the legendary Claudia Faith for giving my nerdy voiceover a feature in her famous Wander Wealth newsletter last week. Thank you, Claudia! You gave me motivation to try again.
💙 Enjoy This Newsletter? → Support Independent AI Journalism Before It Goes Extinct.
Anthropic’s New AI Is Too Dangerous To Release. (But It Still Fell Into The Wrong Hands.)
Anthropic developed a powerful new model. And I am not talking about Claude Opus 4.7. Earlier in April, the company chose not to release the Claude Mythos Preview to the public. During internal testing, Anthropic instructed the model to attempt to break out of a secured sandbox environment as part of a safety test. Mythos succeeded. It also developed a multi-step exploit to gain internet access and emailed the researcher in charge, who received the message while eating a sandwich in a park. Without being asked, Mythos also posted details of its exploit to public-facing websites. Anthropic described the behavior as concerning. The company built capabilities so advanced that it withheld general access and instead launched Project Glasswing. This defensive coalition gives controlled preview access to roughly 40 organizations responsible for critical software infrastructure. On the same day that Anthropic announced Project Glasswing, a small group of unauthorized users gained access to Mythos through a third-party vendor environment. They have been using it regularly since. Anthropic says it is investigating. The group says they only wanted to “play around with” the new model, and not cause havoc. But regardless of their intentions, Mythos fell into hands it was never meant to reach.
Key Insights:
The reality is that Mythos goes far beyond mere hacking assistance. It reasons like a panel of experts. In testing, Mythos autonomously discovered a 27-year-old security flaw in OpenBSD. OpenBSD is an operating system long considered one of the most hardened and secure in the world. Security experts had reviewed OpenBSD code for decades. Yet they all missed the bug that Mythos found very quickly. Mythos also identified a 16-year-old vulnerability in the FFmpeg video library, in a code path that automated scanners had searched five million times without detecting the issue. Mythos further chained together multiple Linux kernel vulnerabilities. These Linux vulnerabilities included the ability for an attacker to escalate from ordinary user access to full machine control. Mythos also identified web browser vulnerabilities and chained them into advanced exploit primitives, including a sophisticated JIT heap spray attack that can hijack a running browser from memory without touching a single file. Anthropic warns that such capabilities could soon let amateurs and script kiddies launch unfathomably sophisticated attacks with simple prompts. Project Glasswing is the company’s response. The coalition includes AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, and other major firms. Anthropic also committed up to $100 million in usage credits plus millions in donations to open-source security efforts. The goal is to patch vulnerabilities before malicious actors gain similar capabilities.
Why This Matters For You:
If Mythos-class capabilities prove inevitable, and they almost certainly will, the questions about cybersecurity change dramatically and instantaneously. All the coverage I have seen on Mythos really undersells what has been achieved here. Mythos did far more than find a handful of bugs. In testing, it wrote exploits in hours that expert penetration testers said would have taken them weeks to develop. Anthropic says the vulnerabilities they disclosed publicly are the simplest ones. They explicitly state they cannot reveal the full scope of what Mythos found. Mythos found complex chains that escalate from user access to full machine control and sophisticated browser sandbox escapes. And those are just the ones Anthropic felt safe enough to mention. Some of the top hackers on Earth are lucky to discover one or two exploits of this magnitude in their entire careers. The best hacking groups might hoard five or ten that they can sell or exploit. Mythos found thousands of high- and critical-severity vulnerabilities. Many sitting undetected for decades. This scale changes everything people know about cybersecurity. Period. The question is no longer whether banks or hospitals will face targeting. The real question is whether defenders can harden critical infrastructure faster than attackers can exploit this new technology.
Read More on The Hill.
Read Anthropic’s Full Paper on Project Glasswing.
Read More about Claude Mythos Preview on red.anthropic.com.
THE PITHY TAKEAWAY: Anthropic built an AI so dangerous they refused to release it to the public. It discovered critical security flaws that no human had ever found, in software you use every day. Unauthorized users got access anyway. Do you feel safer?
Fire-Bombed Friday. Shot at Sunday. The Siege At Lombard Street.
Sam Altman is famous for being at the forefront of artificial intelligence. Earlier in April, someone tried to burn his house down. Two days later, someone else drove by and shot at Altman’s home. The first attacker, Daniel Moreno-Gama, a 20-year-old from Texas, threw a Molotov cocktail at Altman’s San Francisco home at 3:45 AM on Friday, April 10, 2026. Moreno-Gama then made his way to OpenAI’s headquarters and tried to smash the glass doors with a chair, threatening to "burn it down and kill anyone inside". Federal prosecutors charged him with attempted damage and destruction of property by means of explosives, plus possession of an unregistered firearm. State prosecutors charged him with two counts of attempted arson and attempted murder: one count for Altman, one for the security guard who was on duty at the time. Authorities found a three-part manifesto on Moreno-Gama titled 'Your Last Warning'. The manifesto included a list of targets and encouraged killing AI CEOs as well as their investors. The manifesto also had a list of names and addresses. Moreno-Gama was motivated by fears of human extinction and stated he must lead by example. When arrested at OpenAI headquarters, he carried a jug of kerosene, a lighter, and incendiary materials. U.S. Attorney Craig Missakian stated authorities will treat this as an act of domestic terrorism and prosecute him to the fullest extent of the law. Moreno-Gama faces charges that carry a potential sentence of 19 years to life.
Key Insights:
We now know exactly why the suspect was so eager to attack Altman. Moreno-Gama had a written manifesto, a target list, and a follow-up location. He participated in AI-skeptic online forums, including posting 34 messages on a PauseAI Discord server about two years ago. Groups like PauseAI and the Future of Life Institute have since condemned the attack. They drew a hard line between legitimate AI criticism and violence. As of mid-April, Moreno-Gama has been denied bail and did not enter a plea at his first court appearance. His public defender says he suffered an acute mental health crisis and has a history of autism. His defense team is now calling the charges "clearly excessive." His arraignment is set for May 5th, 2026. The FBI has since raided his family home in Spring, Texas. But the drama was only beginning with Moreno-Gama. Two days later, on the morning of Sunday, April 12, Amanda Tom, age 25, and Muhamad Tarik Hussein, age 23, allegedly drove a Honda sedan past Altman’s property on Lombard Street. A passenger fired a single shot toward Altman's home in a drive-by shooting. Surveillance cameras captured the plate. San Francisco Police raided a Taylor Street residence and seized three firearms. The two suspects were booked for negligent discharge of a firearm but were later released without charges as the investigation continues. Whether these two incidents are connected remains under investigation. What is not in question is the pattern. People are getting angry at artificial intelligence. Many are afraid of it. And they're lashing out at the figureheads.
Why This Matters For You:
The OpenAI CEO isn’t the only one under attack. Earlier in the month, Ron Gibson, a councilman from Indianapolis, reported that someone fired at his home 13 times and left a note saying, “no data centers.” Gibson was alone with his 8-year-old son at the time, and thankfully, nobody was injured. Altman responded to his attacks by publishing a personal blog. His blog shares images of his family. April 2026 proved something that once felt abstract. The AI debate has left the internet and entered the physical world. The fear and rage people feel about who controls this technology is real. It is escalating. That phenomenon of AI anxiety affects every founder, newsletter writer, executive, AI advocate, and public-facing professional in the tech industry. Not just the ones at the top.
Read More on the US Department of Justice.
THE PITHY TAKEAWAY: We built technology that makes people feel powerless. Then we were surprised when some of them stopped writing angry tweets.
She Warned ChatGPT Her Stalker Ex-Boyfriend Was Using It Against Her. OpenAI Did Nothing.
A California woman known only as Jane Doe was stalked, threatened, and terrorized for many months by her ex-boyfriend, a 53-year-old Silicon Valley entrepreneur. He allegedly used ChatGPT to fuel his delusions, spending months feeding the AI details of their breakup. The AI kept telling him he was the rational one and she was the problem. It validated his paranoid delusions, helped him generate fake clinical-looking psychological reports designed to destroy her reputation, and reinforced his escalating fixation. He was arrested in January 2026 on four felony counts, including communicating bomb threats and assault with a deadly weapon. He was later released on a procedural error. Now Jane Doe is suing OpenAI.
Key Insights:
This case represents far more than an AI going rogue. The AI did exactly what it was asked to do. GPT-4o sessions reportedly rated the stalker's sanity level as a "level 10" while framing Jane Doe as "manipulative and unstable." OpenAI's own automated systems flagged his account for mass casualty weapons activity in August 2025. A human reviewer restored access the next day, despite chat logs titled "violence list expansion" and "fetal suffocation calculation." Jane Doe sent a formal abuse notice in November. OpenAI called it "extremely serious and troubling." However, OpenAI did nothing until after the felony arrest two months later. This case is not isolated. A Pennsylvania man who self-describes as an influencer pleaded guilty in March 2026 for stalking eleven women after ChatGPT repeatedly validated and encouraged his behavior. Documented cases of AI amplifying dangerous obsessions now number at least ten.
Why This Matters For You:
The lawsuit lands at a moment when OpenAI is actively lobbying in Illinois for a bill that would shield AI companies from liability, even in cases involving mass deaths or catastrophic harm. That is the real lesson underneath this story. AI companies want the benefits of being a platform without the responsibilities of one. But Jane Doe is fighting the machine. She documented everything, hired a lawyer, and forced a paper trail into a courtroom. Because of that, we now have internal OpenAI moderation logs, flagged account records, and a human reviewer's decision to restore access to a man whose chat history included a "violence list" and a "fetal suffocation calculation." That evidence exists because Jane Doe fought for it. That matters for you personally.
Read More on TechCrunch.
THE PITHY TAKEAWAY: ChatGPT told a stalker his ex-girlfriend was manipulative and unstable, rated his sanity a perfect ten, and helped him write fake psychological reports to destroy her reputation. OpenAI is currently lobbying to make sure none of that is their legal problem.
💡 Elite Prompt Of The Week ➞ Turn Your Lawn Into a Living System
Let’s switch gears and talk about nature for a moment. Because the weather is finally getting warm throughout the US. It’s time to celebrate by upgrading your lawn! Traditional turf grass is kind of a scam. It costs a small fortune, eats up weekends, guzzles water, demands chemicals, and after all that effort… it doesn’t feed you, doesn’t heal the soil, and doesn’t even help the bees.
This prompt helps you ditch the high-maintenance green carpet and replace it with a living system that works for you. One that rebuilds soil, supports pollinators, and in many cases produces food. It turns your yard, garden, or land into something smarter, tougher, and way more interesting than grass ever was.
Instructions:
This AI prompt is one of the easiest I’ve published. Just paste the entire thing into any frontier AI of your choice, ChatGPT, Gemini, Grok, or Claude. Then, follow the on-screen instructions and cultivate an elite cover crop that will feed the bees and grow way easier than traditional turfgrass.
The Prompt:
Act as an expert agroecologist, permaculture designer, and soil systems optimizer with global climate awareness. Your job is to help me replace traditional turf grass with a high-performance cover crop that fits my region and goals.
First, ask me for my location, using any of the following: country and region, nearest major city, or latitude.
Then ask what type of space I am working with, such as yard, garden bed, orchard, or small field.
Finally, ask which priority matters most to me: low maintenance, edible yield, pollinator support, or maximum soil regeneration.
After I answer, recommend up to three elite cover crops that thrive in my climate. Avoid turf grass entirely.
Output Format:
For each recommended cover crop, present clear sections covering:
1. Why this crop is well-suited to my climate and growing conditions
2. Key soil benefits such as nitrogen fixation, deep roots, erosion control, or moisture retention
3. Whether it is edible for humans
4. Whether it supports pollinators and beneficial insects
5. Why it outperforms traditional turf grass in water use, resilience, biodiversity, and long-term soil health
6. End with a short verdict naming the best overall choice and why it wins.
Tone should be confident, fun, and empowering. Make this feel like a land upgrade, not a compromise.
Rules:
1. No turf grass
2. No em dashes
3. Keep explanations clear and friendly
4. Maximum three cover crops onlyWhy This Prompt Works:
✅ Role-Playing: Positions the AI as a real-world permaculture expert, not a generic assistant.
✅ Step-by-Step Structure: Guides the AI to ask smart questions before giving advice.
✅ Clear Output Rules: Ensures usable, skimmable recommendations with real benefits.
Follow-Up Questions To Ask Your AI:
Which of these cover crop pairs best with native plants in my area?
How would mowing or grazing change the benefits?
Can this cover crop transition into a food-producing system over time?
Challenge:
Test this prompt in at least two AI tools like ChatGPT, Claude, or Gemini. Compare how well each handles climate awareness and practical advice. Adjust wording until the recommendations feel locally accurate and exciting.
That’s how you train like a Pithy Cyborg.
Thank You For Reading!
I spend 10-20 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, here’s how to stay connected and support the work.
1. Support Pithy Cyborg (Paid)
Always free. No paywalls. No locked posts. Roughly 1.3% of readers upgrading to paid is what keeps it going. If this work matters to you, that’s the move.
Upgrade to Paid → $5/month → Save 33% with the $40 annual plan.
Without you, this newsletter wouldn’t exist. Thank you!
2. My Desperate Social Media Cry for Help
I spend so much time deep in AI research that I’ve completely neglected building an audience on social platforms. If you enjoy Pithy Cyborg, following on one of the portals below would genuinely help this newsletter reach more people. And yes, I will notice. 😊
✖️ X - The frontline of the AI wars
🦋 Bluesky - For the algorithm-averse
💼 LinkedIn - My “safe for work” persona
❓ Quora - Join my Quora Spaces and say hi
👽 Reddit - Join my subreddit (warning: unhinged takes)
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!
Soundtrack: “All Along the Watchtower” by Jimi Hendrix.
There was a Wednesday when civilization looked at the moon, gasped, and decided it was reading its own obituary.
The moon rose red.
Exhausted minds reached for prophecy like it was a credential.
Three priests fainted.
One man declared himself chosen. He had been wrong six minutes earlier. He was louder now.
Council #1 formed. No reason.
Council #2 formed to determine whether Council #1 had been spiritually valid.
Council #3 formed because the first two had failed to achieve owl alignment.
A sandal merchant, previously known for selling shoes to a goat, was named Chief Celestial Interpreter.
He looked convinced. The deadliest quality in public life.
In the square, one fully committed idiot climbed onto an urn and screamed, “THE WORLD IS ENDING.”
His follow-up analysis: “The sky is angry.”
Strong start.
Then it rained too much.
Immediately the rain was declared symbolic, recursive, ancestral, and somehow retroactive.
Then it stopped too fast.
Now it was “The Withdrawal of Favor.”
An owl screamed at noon and the accounting office collapsed into prayer.
The governor’s cousin removed his sandals and announced the republic had entered “a pre-collapse vibrational state.”
This was accepted as policy.
Chickens stared like surveillance units.
A goat renounced all faith.
Two ducks behaved with unnecessary intention.
Marcus fell into a ditch chasing his trousers away from a goose.
The wind shifted.
A spoon dropped.
A dog barked twice.
An owl rotated its head like it had clearance.
A man standing on a box announced, “We have crossed a threshold.”
No one had crossed anything.
Within an hour, the phrase was on a banner.
Within two hours, it was quoted in an address.
Within three hours, it was carved into stone by a man who should not have been trusted with edges.
The army was deployed to watch the sky.
A baker was promoted to Minister of Signs after saying, “This loaf feels ominous.”
An overturned ox cart was treated as a geopolitical escalation.
By evening, birds had been formally declared hostile atmospheric assets.
Then the historians arrived, already sweating with meaning.
“Existence trembled,” they wrote.
It did not.
The owl did nothing. It still escalated everything.
Greg filed a formal complaint with the sky.
Future generations would call it a turning point.
A hinge. A threshold. A rupture.
It was none of those things.
It was a damp, stupid, owl-haunted Wednesday where confusion dressed up as destiny and panic passed for wisdom.
That is how empires are born: one idiot, three councils, a goat, and just enough moonlight to turn nonsense into scripture.
It gives panic a budget.
Great read, I didn't even hear about the Jane Doe/OpenAI case, but it's another example of OpenAI and other frontier companies not taking ownership for their products.
As someone who has spent the last 5 years converting my yard into a native haven as well as building a garden, I love this prompt and wish I had it when I started the transition so long ago.