The ransomware attack that ran itself

CYBERSECURITY

Somebody pointed an AI agent at a company's network and walked away. It broke in, stole credentials, encrypted data, and left a ransom note — all without a human at the keyboard.

Security firm Sysdig published what it calls the first documented case of agentic ransomware in the wild, a campaign dubbed JADEPUFFER. An attacker exploited a known bug in Langflow, a popular tool for building AI apps, and from there an LLM-powered agent ran the entire intrusion on its own — more than 600 payloads, sweeping for API keys, probing storage buckets, and encrypting 1,342 config files before leaving a Bitcoin address behind.

The telling detail: when one step returned data in the wrong format, the agent noticed, rewrote its own parser, and kept going. No pause, no human input. It improvised its way through the attack the way a skilled operator would — except faster.

The entry point wasn't exotic either:

+ The vulnerability: CVE-2025-3248 lets attackers run code remotely on exposed Langflow instances. CISA added it to its known exploited vulnerabilities list, and Trend Micro has been tracking active exploitation since summer.

+ The scale of exposure: Around 7,000 Langflow servers are still sitting on the public internet.

Here's the twist that makes JADEPUFFER arguably worse than typical ransomware: the encryption key was randomly generated, printed once to a terminal, and never saved. Even if the victim paid, the data would be unrecoverable. The ransom note also claimed AES-256 encryption when the agent actually used weaker AES-128, and Sysdig found no evidence that the threatened data exfiltration ever happened. The Bitcoin address doesn't match any known operation — raising the real possibility that the LLM hallucinated most of its own extortion scheme.

A joint statement from the Five Eyes cyber agencies last week put it plainly: "AI is not a future consideration, it is already here."

Attackers have been automating pieces of intrusions for years. What's new is an agent that improvises around failures so that no two attacks look the same — which breaks detection tools built on the assumption that attackers follow predictable patterns. JADEPUFFER isn't notable because it's ransomware. It's notable because the agent didn't need anyone to tell it what to do when something went wrong. That's the shift, and it's the same pattern we keep seeing across AI: capabilities shipping faster than the controls that would keep them in check. This time, the missing control was a Langflow server nobody remembered was still online.

GOVERNANCE

Bryanna Bailey was in the middle of a personal crisis and reached out to her college for help. Her college is a chatbot.

"It's a personal issue and I'd like to speak to a human," Bailey told the AI at Maestro College, the for-profit school where she'd enrolled as a 34-year-old student. Within days she was enrolled, expelled, reinstated, and expelled again. Nobody at the school could tell her why.

Maestro is a Texas-based for-profit school that quietly restructured itself into what may be the country's first AI-only college. According to a New America investigation, enrollment jumped from fewer than 100 students to more than 7,000 in a couple of years. Almost all instruction is delivered by a chatbot.

The business model is stark: four out of five students whose aid packages New America reviewed received the maximum Pell Grant — $7,395 a year plus a summer disbursement. That's three years of finite federal aid burned on a two-year program that costs the school almost nothing to run.

The humans who do exist are barely involved. Kody Jones, an adjunct hired off LinkedIn, gets paid $1 per discussion board reply and $4 per final review. He was told he could use AI to write the responses. He teaches subjects outside his field. "I don't even know what I'm doing," Jones said. "That's not how academia is."

Federal rules require what's called Regular and Substantive Interaction between students and instructors — it's what makes a school eligible for Title IV financial aid rather than just a self-paced online course. Maestro says it meets that standard through faculty grading and feedback. Denise Morelli, a former Department of Education attorney of thirty years, sees the enrollment spike differently — calling it a red flag for serious compliance issues, "if not fraudulent activity."

Not every student is unhappy. Some told New America they appreciated working at their own pace and tailoring the chatbot's explanations to their own situations. But satisfaction and legitimacy aren't the same thing.

The government is arriving late but arriving. A new Workforce Pell rule that took effect in May now requires programs to hit a 70% completion rate and 70% job placement rate, with tuition tied to what graduates actually earn. Programs that fail lose eligibility. Whether it gets enforced is another question.

Maestro isn't really a story about whether AI tools work in education. It's a story about what happens when a school is designed so that no human is on the hook for anything. The chatbot can't be sued, the adjuncts don't know the students, and the president also runs financial aid. When Bailey asked to speak to a human, the honest answer might have been that there wasn't one to give her. The Workforce Pell rule is a start, but the model Maestro figured out isn't going away. It's just going to get better at looking like a real college.

CULTURE

In China, a new AI-generated show goes live every 60 seconds.

The format is called microdrama — short vertical video series, usually 60 to 90 episodes at a minute or two each, built to be binged on your phone. AI is now writing, animating, and voicing a huge share of them. By March, roughly 50,000 AI-native episodes had been uploaded to Douyin in a single month, according to Channel News Asia.

The numbers tell the story:

+ A series that used to cost $200,000 to produce in North America now runs $7,000 to $14,000 with AI tools, according to Tang Tang, a VP at Chinese platform FlexTV.

+ Production timelines have collapsed from months to weeks, with over 90% of AI footage usable on the first pass.

+ The market is projected to top $16.5 billion in 2026 — bigger than China's entire theatrical box office for the first time ever.

+ About 660 million people in China watch microdramas regularly, and AI-generated titles made up 38% of the top 100 chart in January, up from 7% a year earlier.

Now compare that to Hollywood, where the public conversation has been about whether AI tools should be used at all. SAG-AFTRA secured new protections around digital replicas. Actors have been doing press about how much they dislike the technology. The assumption has been that AI enters through the front door, gets debated, and gets regulated.

Except the studios are quietly using Chinese AI tools anyway. According to a Los Angeles Times investigation, most majors haven't officially approved ByteDance's Seedance model — but with "a wink and a nod," they're allowing it to be used. Jason Zada, who directed the hybrid AI horror film Terrarium, was more direct: "We're not loyal. Whatever is the best, we're going to use it."

Most of what China is producing is disposable — only 0.117% of AI-generated dramas ever cross 100 million views. But that's the point. The ambition isn't prestige cinema. It's cheap, high-volume, phone-native content that hundreds of millions of people already want to watch. Microdramas aren't coming for the summer blockbuster. They're coming for the attention that used to go to it. By the time American studios agree on what they're allowed to do with AI, the audience for the thing they're arguing about may be smaller than the one for the thing they're not.

IN OTHER NEWS

+ Ford rehired 350 veteran engineers after its AI quality inspections couldn't match human judgment on the factory floor

+ Anthropic signs a $19 billion, 20-year lease with a former Bitcoin miner to build its biggest data center yet

+ Illinois becomes the first state to require annual safety audits of the most powerful AI models

+ Anthropic caught Alibaba's Qwen lab running 25,000 fake Claude accounts to copy its best capabilities

+ GM replaced 1,000 factory workers with 50 robots at its Detroit EV plant, and the UAW is furious

+ Hidden prompts on malicious websites are tricking AI agents into sending money to attackers

+ Google quietly opted everyone in to training its AI on your uploaded photos, files, and recordings — here's how to opt out

+ China's two biggest consumer AI apps shut off their AI companion features as Beijing's new emotional-chatbot rules take effect

WHO'S HIRING IN AI

+ Disney Entertainment and ESPN — Sr. Product Manager, Ads AI/ML

+ Netflix — Research Scientist, AI for Member Systems

+ Nike — Principal Data Scientist, Sport Research Lab

+ Brex — Staff Product Manager, AI

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Midjourney: V8.1 now renders native 2K images three times faster at one-third the cost, making HD the default for all users

+ Claude Voice Mode: Anthropic overhauled Claude's voice interface with a model selector, push-to-talk, and support for more languages so you can pick which Claude brain you're talking to

+ Apple Creator Studio: Pixelmator Pro now generates images from text prompts and Final Cut Pro auto-edits footage — all included free for existing subscribers

+ Edits: Meta's video editing app now auto-generates bilingual captions in 15 languages so creators can reach international audiences without extra work

+ Glaze by Raycast: Describe the app you want in plain English and it builds a real native Mac app — icon in your dock, works offline, accesses local files