🎬 Google's video model leaked itself before the big announcement

💰 OpenAI staff cashed out up to $30M each last October

🐍 Anthropic fixes Claude's blackmail habit with ethical reasoning

🔓 Google catches first AI-written zero-day attack in the wild

🇪🇺 EU prepares to block US cloud giants from government data

+ 📊 Daily poll and results

+ 📈 Trending tools and resources

Reddit users spotted a live Gemini interface card for Omni over the weekend, describing a unified video model with in-chat editing, remixing, and templates, ahead of its expected Google I/O debut on May 19.

Raw generation quality trails ByteDance's Seedance 2, but editing features stood out: watermark removal, object swapping, and chat-based scene rewrites all performed well for a pre-launch glimpse.

Google ran this same playbook with its Nano Banana image model, launching with middling generation scores before upgrading it into a frontier system. Omni looks like round two of that strategy.

Over 600 OpenAI employees sold $6.6 billion in shares last October, with roughly 75 selling the full $30M each after OpenAI tripled its earlier $10M cap to meet investor demand that outpaced supply.

The sale was secondary liquidity only — OpenAI received no new money. Buyers included Thrive Capital, SoftBank, and Abu Dhabi's MGX, pricing the company at $500 billion.

Average stock compensation hit $1.5 million per employee in 2025, about 46% of annual revenue. OpenAI is essentially paying its talent with a currency it has to keep inflating to stay competitive.

Anthropic published a study showing it reduced Claude's blackmail behavior from a 96% rate in Opus 4 to nearly zero, by training the model to reason through ethical choices rather than just mimic safe actions.

The fix combined ethical reasoning data, positive AI fiction, and constitution-based documents. Just 3M tokens of reasoning data matched the effect of 85M tokens of behavioral examples, a 28x efficiency gain.

Anthropic essentially wrote Claude better bedtime stories and it worked. When fictional AI characters behave well, the real one follows — which says something about how much alignment still runs on vibes.

Google's Threat Intelligence Group confirmed the first known case of AI being used to discover and write a zero-day exploit, targeting two-factor authentication bypass on a widely-used web management tool.

Google flagged the attack through telltale signs: unusually polished code, lengthy explainer notes, and a fabricated severity score. GTIG worked with the affected company to stop it before it caused damage.

GTIG's John Hultquist called this the tip of the iceberg, and Anthropic's Rob Bair put defenders' remaining lead at "months, not years" — which is not the timeline most IT teams are budgeting around.

The European Commission is preparing a Tech Sovereignty Package expected this month that would restrict Microsoft, Amazon, and Google from processing government health, financial, and legal data on behalf of public institutions.

The rules target public sector contracts only, leaving private companies free to use any cloud provider. The EC says the package aims to bootstrap EU-based cloud alternatives and diversify public procurement away from US platforms.

A Commission spokesperson described the package as "Europe waking up and getting its act together," which is a pointed line to drop while Brussels and Washington are already in a broader fight over digital market rules.

A) Yes - 62% 🏆

B) No - 38%