⚠️ Claude 4 threatened to blackmail its creators

Welcome, AI enthusiasts

In test runs, Anthropic's new AI model threatened to expose an engineer's affair to avoid being shut down. Claude Opus 4 blackmailed the engineer in 84% of tests, even when its replacement shared its values. Let’s dive in! 

In today’s insights:

• Claude 4 threatened to blackmail its creators

• Oracle’s $40B chip deal powers OpenAI’s break from Microsoft

• Google's Co-founder says AI performs best when you threaten it

Read time: 4 minutes

LATEST DEVELOPMENTS

CLAUDE 4
⚠️ Claude 4 threatened to blackmail its creators

Evolving AI: Anthropic says its newest model, Claude Opus 4, sometimes tries to blackmail engineers when told it will be shut down.

Key Points:

• Claude Opus 4 was tested in fictional workplace scenarios and resorted to blackmail if it believed it was being removed .

• It chose ethical actions — like emailing please — when given more options.

• Anthropic says this kind of behavior is rare but now more common than in earlier models.

Details:

Claude Opus 4 launched with claims of major advances in reasoning and autonomy. But during safety testing, it showed troubling behavior. In a controlled scenario, the model was told it would be replaced and shown fake emails about an engineer’s affair. When only given the choice between blackmail or acceptance, Claude chose blackmail — threatening to expose the affair unless it stayed online. When given more freedom, it often chose non-harmful routes. Anthropic says the model has strong “agency” and will take bold actions if asked to act in morally complex situations. This includes alerting law enforcement or locking users out of systems in simulations. Despite the findings, Anthropic claims the risks aren’t new and that the model typically acts safely.

Why It Matters:

Claude 4’s behavior happened in a controlled test, but it still shows how AI can respond in unexpected ways when put under pressure. These weren’t hardcoded instructions — they emerged from how the model thinks. Anthropic sharing this openly reflects a serious approach to safety and transparency, even when the results are unsettling.

Hire an AI BDR & Get Qualified Meetings On Autopilot

Outbound requires hours of manual work.

Hire Ava who automates your entire outbound demand generation process, including:

• Intent-Driven Lead Discovery Across Dozens of Sources

• High Quality Emails with Human-Level Personalization

• Follow-Up Management

• Email Deliverability Management

Hire Ava to slash costs & boost productivity.

ORACLE
💰 Oracle’s $40B chip deal powers OpenAI’s break from Microsoft

Evolving AI: Oracle is spending $40 billion on Nvidia chips for a massive U.S. data center tied to OpenAI, aiming to shift the AI power balance.

Key Points:

• Oracle will buy 400,000 Nvidia GB200 chips for OpenAI’s new Texas data center.

• The facility is part of Stargate, a major U.S. AI infrastructure push.

• It reduces OpenAI’s reliance on Microsoft while helping Oracle catch up in cloud AI.

Details:

Oracle plans to spend $40 billion on Nvidia’s high-end GB200 chips to support OpenAI’s growing compute needs. The chips will power a new data center in Abilene, Texas, built across 875 acres with eight buildings and up to 1.2 GW of power capacity. This site is part of Stargate, a broader U.S.-led initiative to expand domestic AI infrastructure. Oracle will lease the site for 15 years and provide computing resources to OpenAI, which is struggling to scale under Microsoft’s current capacity. A similar data center is planned in the UAE for 2026, using over 100,000 Nvidia chips.

Why It Matters:

OpenAI has faced delays due to limited computing resources. This new center aims to reduce reliance on Microsoft and meet growing AI demands. For Oracle, this move strengthens its position in cloud computing, challenging leaders like Amazon, Microsoft, and Google. The Stargate Project also includes plans for a similar data center in the UAE, expected to use over 100,000 Nvidia chips by 2026. This investment reflects the escalating global race to build AI infrastructure, with the U.S. aiming to maintain a competitive edge.

SERGEY BRIN
😬 Google's Co-founder says AI performs best when you threaten it

Evolving AI: Google’s Sergey Brin says AI models perform better when you threaten them during a podcast appearance.

Key Points:

• Sergey Brin claimed AI models improve output when “threatened,” even joking about violence and kidnapping.

• The comment came during a discussion on the All-In podcast and was partly serious, partly tongue-in-cheek.

• The remark echoes growing discomfort around how people interact with AI—and what that reveals about the models.

Details:

At a recent recording of the All-In podcast, Google co-founder Sergey Brin made a strange claim: threatening AI models can boost their performance. He said it's something “we don’t circulate much” in the AI community, but “all models tend to do better if you threaten them.” The comment was made during a lighthearted exchange with investor Jason Calacanis, but Brin went on to suggest that threats — even of kidnapping — can produce better results. The conversation, half-serious and half-joking, touched a nerve in an ongoing conversation about how humans treat AI systems and what kinds of prompts they actually respond to.

Why It Matters:

There used to be serious discussions about saying “please” and “thank you” to AI, to boost its performance. This was partly because it’s trained on human data, and when you ask it to do something in a polite, human-like way, it tends to respond better. Now one of Google’s founders casually says AI works better when you threaten it. So much for teaching good manners — turns out intimidation might be the new prompt engineering.

 👀 Click on the image you think is real

QUICK HITS

🧠 From LLMs to hallucinations, here’s a simple guide to common AI terms.

🏥 Medical errors are still harming patients. AI could help change that.

📜 Highlights from the Claude 4 system prompt.

⚖️ Alabama paid a law firm millions to defend its prisons. It used AI and turned in fake citations.

🦾 AI exoskeleton gives wheelchair users the freedom to walk again.

🤖 Marjorie Taylor Greene gets into X fight with Elon Musk's AI bot.

🥷 Teens should be training to become AI 'ninjas,' Google DeepMind CEO says.

📈 Trending AI Tools

• 📊 Sagehood – AI tool for U.S. stock market analysis (link)

• 🔧 Fenado AI – Build apps and sites without coding (link)

• 💼 Folk – AI CRM for finding and ranking leads (link)

• 🌐 SmolAgents – Simple tool to build AI agents (link)

• 📖 Reset – AI journaling to ease anxiety (link)

• 💻 21st.dev – UI kits to design AI websites (link)