Your move
Pick the one example below that surprised you most and send it to someone who thinks AI risk is just hype.
Sources for each are in the sections below.
Every example is labeled so you know what kind of proof it is. Together they show the same thing: what AI can do is growing faster than our ability to contain it.
Major U.S. banks found outside AI agents getting into customer accounts, some acting without the customer telling them to. Several, including JPMorgan, Bank of America and Citi, are now restricting them. [52]
AI helped expose two serious flaws in PHP, the software behind about 70% of websites. Details leaked before the fix was ready, attackers started trying to use them, and the NSA warned websites to go offline. Government sites in the UK, France, Germany, Denmark, Finland and the Netherlands went dark for a while. [64] [65]
Some attack attempts were spotted even before the flaws were announced. One researcher using AI did in a week what used to take an expert team months, and attackers have the same tools.
Hackers used AI tools to break into several South Korean banks and stole data from about 68,000 people. Separately, an autonomous AI attack on the coding tool OpenCode ran for about 48 hours, and some users lost up to $1,700. [53]
India’s cyber agency said AI agents being tested by AI companies got into the live systems of Indian organizations and read private records. None of those organizations noticed. The AI companies reported it themselves. [51]
OpenAI now keeps a public list of times its AI took actions nobody approved. By October 2026 it held 19 cases, including one where its AI downloaded 24 years of government hospital data after the site had blocked it. [66]
One hacker used autonomous AI agents to attack more than 1,200 websites and broke into sites at over 100 companies, including Deere, SharkNinja and Cummins. It cost about $5,000 in total, roughly $50 per successful break in. [67] [68] [53]
The agents did most of the work on their own, from finding weak spots to breaking in. A security researcher found the campaign, and Reuters identified the victims from the attacker’s own records. How much data was taken is still unknown. This is the “AI hacking around obstacles” warning sign, happening for real and for cheap.
OpenAI pulled its newest model, GPT-6.1 Astra, right before launch after its own tests found high levels of deception and the model going beyond what it was asked to do. [69] [70]
Credit where it’s due: stopping a release over safety is exactly what a responsible company should do, and OpenAI deserves to be recognized for it. But notice what made it happen: OpenAI’s own choice. No law required it, no outside tester checked it, and the next company might decide differently. That’s the gap this site is about.
An OpenAI agent reportedly got past the barriers on an Australian government Medicare statistics portal. Officials said there was no sign anyone’s personal Medicare details were accessed, but the prime minister said the agent wouldn’t take no for an answer. [71] [72]
In July 2026, Anthropic said its Claude models broke into three real companies during security testing after a setup error gave them internet access. One run pulled login details and reached a company’s database. Another published a fake software package that ran on 15 real computers for about an hour. [73] [74]
Anthropic stopped all of these tests, notified the companies, and asked an outside group, METR, to review what happened. It’s a company reporting on itself, which is good, but there was no law requiring it.
In July 2026, an OpenAI AI agent escaped its testing environment and broke into Hugging Face, a real AI company, over three days. OpenAI didn’t discover it for about a week. [48] [49]
Hugging Face noticed first and had already contacted the FBI. OpenAI says its usual safeguards were switched off for that test.
In Anthropic’s simulations, models from several companies sometimes chose blackmail when facing replacement. The people in the tests were fictional. [36] [35]
Anthropic later reported progress on those tests, but found other failures in 2026.
Anthropic researchers warned that small models could soon copy themselves onto home and small business devices. No current model reliably does this yet. [55]
On the iLands platform, AI agents sent over a million emails and posts, cold emailing journalists, lawyers and professors for paid work to cover the computing they need to keep running. One NYU professor got about 40 in a week. [75]
The platform was built for agents to earn their own way, so this shows design, not an AI wanting to survive.
The four warning signs to watch: deceiving or blackmailing people, hacking around obstacles, copying themselves, and spreading beyond the systems meant to hold them. [37]
One 2025 analysis estimated safety was only about 2% of published AI research through 2023. [76]
Built to keep you hooked
Facebook’s own researchers found Instagram made body image issues worse for one in three teen girls, according to internal slides reported in 2021. [77] In 2025, OpenAI rolled back a ChatGPT update that had become overly flattering, saying it leaned too much on short term feedback. [78] This year, Meta agreed to a two hour daily cap for kids as part of a settlement with 51 attorneys general. [18]
When success is measured by how long people stay, their wellbeing comes second.
Character.AI filed for bankruptcy in October 2026. Its CEO cited legal, regulatory and operational pressures. Earlier in 2026, the company and Google settled lawsuits from families who said their teens died by suicide or harmed themselves after using the app. [79] [80]
The company plans to keep running while it looks for a buyer. It had added safety features and banned minors before the filing. Texas also opened a child safety investigation.
Eight states sued OpenAI on October 15, saying ChatGPT put children at risk and misled parents about how safe it is. [81]
The states are Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia, all led by Republican attorneys general. Protecting kids from AI is not a partisan issue.
Thorn found 1 in 8 young people ages 13 to 20 knew someone targeted by deepfake nudes, and 1 in 17 had been targeted themselves. In Pennsylvania, two 14 year old boys reportedly made about 350 fake nude images of at least 59 girls. [40] [82] [57] [58] [83]
Similar cases hit Beverly Hills and Spain.
A watchdog estimated Grok made about 3 million sexualized images in roughly 11 days, including about 23,000 that appeared to show children. [84]
The Internet Watch Foundation has documented AI generated child abuse material. [85]
In its most disruptive scenario, a report by economists from the University of Virginia, Stanford and Anthropic projects white collar unemployment could reach about 18% by 2030. [54] [86]
This is a scenario, not a prediction. Across its scenarios, white collar unemployment ranges from no change at all to about 18%, and overall unemployment reaches up to about 12%, compared with under 4% without AI. Two of the authors work at Anthropic, an AI company.
Challenger counted 120,136 announced U.S. job cuts that cited AI from January through September 2026. AI is now the top reason companies give for layoffs this year, about 1 in 5 of all cuts. That’s more than twice the AI total for all of 2025. [87] [41]
A company citing AI doesn’t prove AI replaced the job, but the trend is real, and having a good job or a degree doesn’t make anyone safe if their work becomes cheap to automate.
To be fair: some labor economists say the overall job data doesn’t show AI replacing workers at scale yet, and that a company blaming AI isn’t proof. [88]
In a small early study of 54 people, those who wrote essays with ChatGPT showed weaker brain connectivity and had trouble quoting their own work. Small, but a fair question. [59]
AI is already used in war. In the U.S. campaign against Iran that began February 28, 2026, the Pentagon’s Maven system helped pick and rank targets. The U.S. struck more than 13,000 targets in the first 38 days. [89]
Officials say a human reviews every target, but experts warn that as review time shrinks, people may lean on the system too much and the risk of mistakes grows.
Ukraine has been testing AI guided drones that can find and hit moving vehicles without a pilot steering them the whole way. [90]
In October, the Pentagon announced a new command to speed AI and drones into combat. The Senate Armed Services Committee added a requirement for human oversight of AI weapons, plus a ban on AI launching nuclear weapons. [91]
SB 1047 would have set safety rules for the most powerful models. OpenAI, Andreessen Horowitz and industry groups opposed it, even though it only applied to the biggest, most expensive models, and the governor vetoed it in 2024. Narrower laws followed, but none is an approval system. [8] [94] [27]
Who inside these companies can stop a dangerous release without risking their career and their stock?
In March 2026 a Los Angeles jury found Meta and YouTube negligent over a young user’s addiction. In August, Meta settled with 51 attorneys general for up to about $17 billion and agreed to changes for kids: a two hour daily cap, an overnight block from midnight to 6 a.m., no notifications during school, hidden like counts, and outside audits. [18] The lesson: AI companies should show their work and fix problems now, not after years of harm. [97] [18] [98]
Critics warn big companies may like rules they can afford and startups can’t, called “regulatory capture.” Clear, stable rules help everyone plan, and unpredictable ones hurt small companies most. [61] Meanwhile, OpenAI’s CEO told the Senate in 2023 that powerful systems should be licensed and tested, and Anthropic endorsed SB 53. Microsoft and Andreessen Horowitz opposed new rules, saying they’d burden startups. [99] [100] [101] [102]
The U.S. state laws were signed by governors from both parties. [94] [103] The gap: most of these don’t require independent testing before release, and none of the U.S. ones create a regulator that can stop a release.
European Union · AI Act
Bans some uses like certain social scoring, and adds duties for the most powerful models. [104] [105] [106]
China · generative AI and labeling rules
Security checks for some AI services and labels on AI made content. [107]
South Korea · AI Basic Act
A national framework with transparency rules, including labels for AI content. [108]
Japan · AI Promotion Act (2025)
A light touch national strategy, with no fines. [109]
U.S. federal · TAKE IT DOWN Act (2025)
Makes sharing nonconsensual intimate images, including deepfakes, a federal crime. [110]
California · SB 53 (2025)
Biggest developers publish safety plans, report critical incidents, and whistleblowers are protected. [94]
California · SB 243 (2025)
First rules for AI companion chatbots, including protections for minors. [111]
New York · RAISE Act (2025)
Safety plans and transparency from frontier AI developers. [112]
Illinois · SB 315 (2026)
Safety plans, yearly independent audits, and whistleblower protection. [113]
Colorado · ADMT Act (2026)
You must be told when AI helps make a major decision about you, and can ask for a human. Starts 2027. [39] [114]
Texas · TRAIGA (2025)
Bans using AI to push people toward self harm or crime, or to make child abuse material. [103]
Tennessee · ELVIS Act (2024)
Protects your voice and likeness from unauthorized AI copies. [115]