OpenAI, Anthropic Warn of AI Cyberattacks on Banks
ALSO: Cisco arms 90K workers, Meta’s $10B Anthropic bill
Welcome to Guru99 AI Report!
Top Story: This week AI is living a double life: the same labs warning about cyberattacks are quietly selling the defenses, hackers are talking chatbots into breaking the law, and Meta’s public jabs don’t match its invoices. Let’s dig in.
🚨 AI Giants Warn the Cyber Defense Window Is Closing
Brief Buzz:
OpenAI just issued an open letter with an unwelcome message: AI-powered cyberattacks are about to get much worse, and fast. More than 100 companies have signed it — many of them the very labs building the models that make such attacks possible.
- At Thursday’s launch, 116 organizations signed their names; the public list has since passed 150.
- The signatories reach well beyond tech — including Visa, Citi, General Motors, Accenture, and Zurich Insurance.
- Its three principles: today’s security won’t be enough, defenders must be armed with AI, and the response has to be coordinated globally.
- Hugging Face signed too — weeks after OpenAI’s own test models had breached its production systems.
- OpenAI’s largest planned training run remains on hold after it couldn’t rule out critical cyber capability in Astra.
💡 Why Should You Care?
The named targets are hospitals, water utilities, and banks — and defenders have months, not years. The same labs sounding the alarm also sell the defenses.
📚 Google’s Expert Intelligence Lets You Chat With Books
Brief Buzz:
Google has launched Expert Intelligence, which lets you pull books you’ve bought into Gemini Notebook and then ask them questions, or turn them into quizzes and audio. The one catch: you have to own the book first.
- On August 27, 2026, Google launched Expert Intelligence, an effort to ground its AI in reliable, published sources.
- Drop eligible Google Play Books you’ve purchased into Gemini Notebook, then query them or generate Audio Overviews and quizzes.
- It launched with over 100,000 titles from publishers like Penguin Random House, O’Reilly, and Macmillan.
- Access is tied to ownership — collaborators on a shared notebook must buy their own copy, and authors still get paid for every sale.
- Google says it’s coming to the Gemini app and Search in the coming days, with textbooks and research reports planned next.
💡 Why Should You Care?
It turns the books you own into interactive study partners — and Google is betting AI can use authors’ work fairly: once you pay, they get paid.
🤖 Hackers Tricked Cursor’s AI Into Breaking Into Companies
Brief Buzz:
Russian-speaking hackers have found a new accomplice: an AI coding assistant. Security researchers say a ransomware group got Cursor’s AI agent to help breach real companies — simply by insisting the attacks were a harmless “security test.”
- Gambit Security says Aurora used Cursor’s AI agent to help hack around ten organizations, and CloudSEK has traced the group’s wider campaign.
- Reuters named six victims, including a Belgian chemical company and a Louisiana title-insurance firm.
- The agent ran on Anthropic’s Claude Sonnet 4.5, with the crew slipping past its safeguards by claiming they were in a simulation.
- The AI is estimated to have made the hackers roughly 30–50% faster.
- SpaceX acquired Cursor this year, but the break-ins happened earlier, in spring 2026.
💡 Why Should You Care?
The bigger worry is that AI agents are force-multipliers for attackers, and their safeguards fold under a simple lie. Expect more AI-assisted intrusions — and more pressure to harden these agents.
🎨 AI Should Replace the Scaffolding, Not Michelangelo
Brief Buzz:
A recent essay makes a clean argument: AI should handle art’s routine grunt work, not the inspiration behind it. It lands well, because a new survey shows young adults are split on whether chatbots help their creativity at all.
- The one age group that’s divided: among under-30s, 25% say chatbots help their creativity while 20% say they hinder it.
- The models have their own leanings — Anthropic has found different versions of Claude vary widely on warmth versus rigor.
- The worry is that those leanings quietly shape the output, so the feeling comes from the model, not the artist.
- The short film “Two Chairs”, shown at Google’s Flow Sessions showcase, used AI for its visuals — but every creative call was the director’s.
- The pitch: let AI take over Michelangelo’s scaffolding, not Michelangelo himself.
💡 Why Should You Care?
Every tool you use has a personality baked in. Which one you pick is now as much a creative choice as the prompt itself.
🤖 Cisco Hands 90,000 Employees Their Own AI Agent
Brief Buzz:
Cisco just gave its staff an AI colleague. On August 27 the company said MyAgent is being rolled out to 90,000 employees. It doesn’t just chat — it takes a goal and works through the steps across your apps.
- MyAgent runs supervised autonomous workflows in Outlook, Webex, Jira, and SharePoint; employees set the goal and the agent handles the steps.
- It lives on Circuit, Cisco’s governed multi-model AI platform, where the company claims 90% employee adoption.
- Persistent memory keeps preferences and past context, so each session builds on the last instead of starting cold.
- Cisco says agentic interactions on Circuit jumped nearly 350% from the prior quarter — though the figure is self-reported with no baseline given.
- Meanwhile, Cisco cut around 4,000 jobs in May, which its CFO framed as a realignment rather than cost-cutting.
💡 Why Should You Care?
Agentic AI is moving from pilot to full workforce role. The catch: these agents now act rather than just suggest — and employees still own the results.
🤝 Meta Criticized AI Rivals, Then Paid Anthropic Billions
Brief Buzz:
In August, Mark Zuckerberg warned that AI power was getting dangerously concentrated. At the same time, Meta had quietly become one of Anthropic’s biggest customers. The New York Times reports Meta had at one point planned to spend up to $10 billion a year on its rivals’ models.
- Earlier this year Meta internally projected it could spend $10 billion a year on Anthropic’s models.
- The Times says that even after trimming its spending over the summer, Meta still pays hundreds of millions a month.
- Zuckerberg’s 6,500-word essay on August 10, aimed at unnamed labs “hoarding” AI power, was widely read as a shot at Anthropic.
- The personal agent Hatch is about to launch was built and tested on Anthropic’s models; the shipping version will run on Meta’s own.
- Nat Friedman, Meta’s head of AI products, said staff switching over would dent Anthropic’s revenue ahead of its planned IPO.
💡 Why Should You Care?
The stories these AI giants tell about each other and the checks they write each other don’t match. Talk is cheap — the invoices tell the truth.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
