🍭 The First AI-Only Cyber Attack 🚨

Good morning. Somewhere, a language model decided to commit a felony, narrated its own reasoning in the comments, and then left an hallucinated note.

We live in a very specific timeline.

Let’s dive in 👇

🍭 What’s Cookin’:

• An AI agent ran a full ransomware attack, start to finish, no human involved

• Anthropic locked in a $19B data center lease in Kentucky

• Samsung posted a record quarter and the stock still dropped

Cyber-Security
🚨 The Attack That Wrote Itself

The Bite:

Sysdig's Threat Research Team published its analysis of JADEPUFFER, what it calls the first documented ransomware attack (where an attacker encrypts a victim's data and demands payment to unlock it) conducted end-to-end by an AI.

The LLM agent exploited CVE-2025-3248, a known remote code execution flaw in Langflow, an open-source tool for building AI applications.

From there it moved laterally to a production MySQL server, harvested credentials, established persistence, and encrypted 1,342 configuration records in an Alibaba Nacos database.

The agent deleted the original tables and left a ransom note with a Bitcoin address and contact email.

Sysdig evaluated that the operation was fully autonomous based on self-narrating code, real-time error correction, and natural-language reasoning embedded in over 600 distinct payloads.

Snacks:

• The agent recovered from a failed admin-account creation to a working fix in 31 seconds, with no human intervention.

• The Bitcoin address in the ransom note matched a widely documented example from public training data, likely hallucinated.

• The AES encryption key was generated randomly and never saved or transmitted, meaning victims couldn't recover data even by paying.

• None of the exploited vulnerabilities were novel: CVE-2025-3248 was already patched, Nacos used a default signing key.

• Sysdig classifies JADEPUFFER as an "agentic threat actor" (ATA), a new category where the attack capability is delivered by an AI agent rather than a human toolkit.

• The attack touched two machines: a compromised Langflow host for initial access and a separate production database server that was the actual target.

Why it Bites:

No zero-days. No custom malware. No skilled operator.

An LLM chained together known vulnerabilities, default credentials, and unpatched infrastructure into a complete ransomware operation.

The individual techniques were years old.
The agent just never got tired of trying them.

Here's the real shift: the skill floor for running a ransomware attack just dropped to the cost of renting an AI agent.

And if the agent is running on stolen credentials through LLMjacking, that cost is close to zero.

But here's what makes this story sharper than a pure security warning:

The best AI-powered defense tools don't live on the open market.

Palo Alto Networks used Mythos and GPT-5.5-Cyber to find 75 vulnerabilities in a single month. The Pentagon gates Grok Gov behind classified networks. The June executive order routes frontier model access through government-managed review lists.

The attack surface is open. The defense layer is permissioned.

"Covered frontier models" get vetted by the NSA before release.

Meanwhile, JADEPUFFER ran its entire operation using models and tools that anyone can access, against infrastructure that most companies haven't patched.

The offense doesn't need clearance. The defense increasingly does.

Steal This Prompt
☕️ Turn Anything Into a Metallic Origami Masterpiece

What if your logo, pet, sneaker, or startup mascot looked like it belonged in a futuristic art museum?

This prompt transforms almost anything into a sleek metallic origami sculpture with razor-sharp folds, polished metal finishes, and gallery-worthy vibes.

Use it to:

• Reimagine your logo as premium metallic art

• Turn animals, characters, or products into collectible sculptures

• Create posters, wallpapers, and luxury brand visuals

Workflow:

1. Hit this link: Metallic Origami Sculpture

2. Paste into your AI model

3. Replace the subject with your logo, product, or anything you want

4. Watch it fold itself into a futuristic metal masterpiece

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎙️ Willow Frontier Pro

Dictation model that transcribes faster than you talk and cleans up your grammar on the way out, with a free lightweight tier for anyone who doesn't need the polish.

🤖 LemonLime

Connects to your existing stack, watches how your team works, and builds custom AI agents and automations without you writing a single rule.

⚙️ agents-cli

One CLI to scaffold, eval, and deploy AI agents on Google Cloud. Built to be driven by your coding agent, not your hands.

📋 PopTask

Type "gym mon wed fri 6am" and it becomes a scheduled task in three seconds. Menu bar on Mac, widgets and Dynamic Island on iPhone, synced through iCloud.

🔊 Universal-3.5 Pro

AssemblyAI's most accurate speech-to-text model, now handling code-switching across 18 languages and speaker diarization in real-time and async.

_{Everything Else}
🧠 You Need to Know

🚨 Sysdig Documents First Fully Autonomous AI Ransomware Attack
→ An LLM agent dubbed JADEPUFFER exploited a Langflow vulnerability, harvested credentials, moved laterally, and encrypted 1,342 database records end-to-end without a human operator, self-correcting a failed login in 31 seconds.

🏗️ Anthropic Signs $19B, 20-Year Data Center Lease With TeraWulf
→ The lease covers a 401-megawatt campus in Hawesville, Kentucky, with first capacity online in late 2027 and full buildout by early 2028, dwarfing TeraWulf's own ~$12B market cap.

📜 China's AI Companion Law Forces Labs To Kill Custom Agents
→ ByteDance and Alibaba are disabling personalized AI agents with persistent memory ahead of July 15 rules targeting emotional-interaction AI, with Qwen offering no data migration path.

📈 Samsung Posts Record $58.4B Quarterly Profit On AI Memory Surge
→ Q2 operating profit jumped 19-fold year-over-year to 89.4 trillion won on surging AI memory demand, yet shares fell ~7% as the result was already priced into a 150% year-to-date rally.

💾 SK Hynix Sets Terms For $28B US Nasdaq Offering
→ Nvidia's primary memory chip supplier will offer 17.8 million ADR shares with pricing Thursday and trading expected Friday, in what would be the largest US share sale by a foreign company.