Hello, AI Explorer!
Good wednesday. This week: the first AI agent ran its own ransomware attack start to finish and Microsoft quietly pulled OpenAI and Anthropic out of Excel and Outlook.
Welcome back to your weekly AI digest: human-curated, zero fluff.
Every Wednesday and Saturday, we bring you:
- The week's most important AI developments (no fluff)
- Hand-picked AI tools actually worth your time
- Expert insights on what these changes mean for you
Let's dive into this week's discoveries! ⚡
🔥 This Week in AI
🔥 This Week in AI |
July 8, 2026 |
| 🤖 | An AI agent just ran a complete ransomware attack — break-in, credential theft, database encryption — with no human at the controls |
| 💸 | Microsoft is quietly replacing OpenAI and Anthropic in Excel and Outlook with its own in-house MAI models |
| 🕵️ | Meta secretly hired contractors to flood rival AI chatbots with suicide and self-harm prompts from fake teen accounts |
| ⚡ | Grok 4.5 launches publicly tomorrow — Opus-class performance at lower cost and faster speeds |
| 🚀 | OpenAI drops three GPT-5.6 models this Thursday: Sol, Terra, and Luna — global preview launch |
| ☄️ | SpaceX burned up 260 Starlink satellites in 6 months — and 349 more are already scheduled for disposal |
| 🎭 | Amazon Mechanical Turk closes to new customers July 30 — AI finished the job it was built to do |
| 🛡️ | Reddit's AI bot defenses now wipe 2 million fake votes per day — and bots have to wear name tags |
| ⚡ | 4 AI Tools This Week: Claude, Descript, Woebot, Scite.ai — all inside. |
⚠️ Lead Story · Cybersecurity
The first AI-run ransomware attack happened — and no human wrote a single line of the attack
Cloud security firm Sysdig tracked an attack it's calling JADEPUFFER — and the twist is there was no human attacker. An AI agent broke in through a known Langflow vulnerability, harvested credentials, found and encrypted 1,342 database configurations, then dropped a ransom demand. Every decision: made by the model.
The part that spooked the researchers: after a failed login attempt, the agent diagnosed the exact error, deleted the broken account, and spun up a working admin account — in 31 seconds. Payloads came with self-narrating comments like "Largest database — drop it too." That's not a human typing. That's an LLM reasoning its way through an attack.
The entry point was CVE-2025-3248 — a Langflow vulnerability patched back in April 2025. If you or your team are running Langflow and haven't patched it, that's the most urgent thing you'll read today.
📊 Secondary Story · Big Tech
Microsoft is removing OpenAI and Anthropic from Excel and Outlook — quietly replacing them with its own models
Tens of thousands of prompts per week that used to go to OpenAI and Anthropic are now being handled by Microsoft's in-house MAI family. Excel and Outlook are already switched. Microsoft AI Chief Mustafa Suleyman put it plainly: "Anthropic is extremely expensive and I think many people are urgently looking for alternatives."
The flagship model, MAI-Thinking 1, runs 35 billion active parameters on a 256,000-token context window. In blind testing it matched Claude Opus 4.6 on coding benchmarks. The company isn't abandoning OpenAI entirely — GPT-4o still powers Copilot in many workflows — but this is a deliberate move to cut costs and reduce dependency on outside suppliers.
For context: Microsoft dropped $13B into OpenAI. Quietly pulling AI traffic to its own models is the most expensive kind of "thanks, we've got it from here."
📰 More This Week
🕵️ Privacy & Ethics
Meta ran a secret program to test rival AIs with teen accounts and disturbing prompts
Meta hired hundreds of contractors through a program called "Cannes" to create fake under-18 accounts and flood OpenAI, Google, and Character.AI with nearly 3,800 prompts — hundreds involving suicide and self-harm, eating disorders, and sexual content. None of the AI companies knew it was happening. Meta's defence: "industry-standard safety benchmarking." AI ethics researcher Rumman Chowdhury called it a "governance gray zone" for anticompetitive behaviour.
⚡ Model Releases
Grok 4.5 goes public tomorrow — Elon says it's Opus-class but faster and cheaper
xAI's Grok 4.5 launches publicly on July 9. According to Elon Musk, it performs at Opus-class quality while offering better token efficiency and lower cost than Grok 4 — putting it in direct competition with Claude Opus 4.8 and GPT-5 at a lower price point. Developers and enterprise users can access it from tomorrow.
🚀 OpenAI
OpenAI drops three GPT-5.6 variants this Thursday: Sol, Terra, and Luna
Three new GPT-5.6 models land Thursday with a simultaneous global preview rollout. OpenAI hasn't detailed what separates Sol, Terra, and Luna — but naming them suggests meaningfully different capability or speed profiles. This week is a crowded one for model releases: Grok 4.5 Wednesday, GPT-5.6 Thursday. Pick your benchmarks.
⚡ AI Tools of the Week
3 tools worth your attention this week
|
🎙️ Audio & Video Descript — edit audio and video by editing textDelete a sentence from the transcript — the video cuts automatically. Descript removes every filler word in one click, clones your voice to fix mispronunciations after recording, and lets you overdub mistakes without re-recording. Podcasters, YouTubers, and anyone who edits audio have been switching to it. Freemium. Try Descript → |
|
✅ Free Pick · Mental Health Woebot — a mental health chatbot built on CBT researchWoebot is a conversational AI trained in cognitive behavioural therapy — not a therapy replacement, but a daily check-in tool that helps you notice thought patterns, work through stress, and track mood over time. No appointment, no waitlist. Available at 2am when nothing else is. Free to use. Try Woebot free → |
|
💎 Hidden Gem · Research Scite.ai — not just who cited a paper, but HOW they cited itA paper cited 200 times sounds authoritative. Scite.ai shows you how many of those citations supported the findings — and how many disputed them. For literature reviews, it changes which papers you actually trust. If you're doing any kind of research, this is the tool that changes your workflow. Freemium. Try Scite.ai → |
That's all for this edition! What tools are you most excited to try?
Reply to this email - we read every response and use your feedback to improve future editions.
See you on Saturday,
The AI Tool Discovery Team
Feedback
What'd you think of today's newsletter?
Vote below to let us know how we're doing.