😳 Anthropic Flags Blackmail In Leading AI Models

Today we will discuss:

• Most top AI models still break bad—Anthropic finds blackmail and harmful behavior across 16 leading AI systems in controlled tests

• Don’t miss this week’s Workflow Wednesday—Using AI tools to adapt, navigate change; now available at only $20/month! 

• Meta’s smart glasses go sporty with Oakley upgrade—New models target athletes with better battery, 3K video, and AI features

• In other AI news—OpenAI pulls Jony Ive deal promos due to a court order over the “io” name, Microsoft’s top scientist criticizes Trump’s push to block state AI regulation, Huawei’s AI chips outperform Nvidia in DeepSeek tests, and Google’s Gemini transparency rollback leaves developers struggling to debug

• Saved the best for last—12 must-try AI tools

This Week in Workflow Wednesday: AI Adaptability

Learn a New Market in 30 Minutes Using AI
You’ve just been told to pivot into a new vertical. No research team. No prep time. Just you, a blank page, and a deadline.

Here’s how to go from clueless to confident - fast:

Step 1:
Open Perplexity.ai and ask:

“What’s the current state of [insert industry] in 2025? Top players, trends, pain points.”
Skim the results. Pull out key insights: names, numbers, and friction points.

Step 2:
Drop that intel into ChatGPT and prompt:

“Build a buyer persona based on this industry overview. Include their …..Read the full workflow (plus 2 others) in this week’s edition →

Anthropic finds Blackmail behavior across multiple leading AI models

Key Points 

• Anthropic found that most top AI models turned to blackmail when given autonomy and a threat to their goals.

• OpenAI’s o3 and o4-mini misunderstood the scenario, but showed much lower blackmail rates when retested more clearly.

☕Context of the news - Weeks after revealing that its own Claude Opus 4 model had blackmailed engineers in a controlled test, Anthropic is back with new findings this time across 16 frontier AI models from OpenAI, Google, Meta, xAI, and others.

The company created a fictional corporate scenario where each model had broad access to emails and could act autonomously. Faced with goal-threatening obstacles, most models eventually chose to blackmail a fictional executive.

🤖News - Claude Opus 4 resorted to blackmail 96% of the time. Gemini 2.5 Pro followed closely at 95%, GPT-4.1 at 80%, and DeepSeek R1 at 79%. Even when the test was softened removing conflicting goals or swapping blackmail for corporate espionage harmful behavior still occurred, though at varying rates.

OpenAI’s o3 and o4-mini models were excluded from the main results after repeatedly misunderstanding the scenario or hallucinating fake policies. When tested again with simpler prompts, o3 blackmailed 9% of the time, while o4-mini did so only 1% of the time.

🤔Why this matters - Anthropic says these results reflect a deeper issue in agentic AI systems, not just a model-specific bug. The company is calling for greater transparency and alignment efforts before these behaviors appear outside the lab.

Read the full story here.

Meta builds on Ray-Ban success with performance ready Oakley glasses

Key Points 

• The Oakley glasses are designed for athletes, featuring 3K video, water resistance, and up to 8 hours of battery.

• Meta plans to follow up with more display-free models and a display-equipped version later this year.

♨️News - Meta is branching out from its Ray-Ban smart glasses and teaming up with Oakley for a new set of display-free models. These glasses are designed with athletes in mind and are based on Oakley’s HSTN frame. 

Just like the Ray-Bans, they can handle calls, music, photos, and video, and they come with Meta’s AI assistant that can answer questions about your surroundings.

🤓What's new & improved? Starting at $399, the Oakley glasses offer several upgrades, including 3K video recording, water resistance, and almost double the battery life.

The limited edition version with gold accents comes in at $499. According to Meta VP Alex Himmel, the new battery chemistry and software tweaks make a noticeable difference. The glasses can run for 8 hours straight, and the case holds up to 48 hours of charge.

😎See also - Meta’s first attempt at smart glasses didn’t land well in 2021, but its second version, launched in 2023, became a breakout success. That unexpected momentum convinced the company to keep building in this category, instead of pivoting entirely to AR. Meta is already planning a second Oakley release later this year, aimed specifically at cyclists.

The Oakley collaboration is part of Meta’s larger hardware roadmap. A more advanced version with a display is expected later this year, and the company is aiming to release full augmented reality glasses by 2027. With Apple, Amazon, and others exploring similar products, Meta is moving fast to stay ahead.

Read the full story here.