Good morning. AI agents might be the future, but today they’re still like that friend who promises to fix your computer and accidentally deletes everything. Just ask Meta’s Director of AI Safety, Summer Yue. She told her OpenClaw agent to scan her inbox and suggest emails to delete without actually deleting anything. Unfortunately, OpenClaw missed the “without deleting” part and enthusiastically began wiping her entire inbox. Yue had to sprint to her computer to stop the carnage. Rookie mistake.
MIT researchers figured out how to exploit chatbot biases. They built a tool that spots hidden biases, moods, and other weird ideas lurking deep inside big AI models like ChatGPT. Instead of guessing, their method pinpoints exactly how these hidden traits shape chatbot responses – and even lets them dial it up or down like a volume knob. The researchers tested over 500 biases, from silly ones like fear of marriage to oddly specific obsessions like being a huge fan of Boston. To prove how easily biases warp chatbot behavior, they made one chatbot rant about NASA faking Apollo 17 photos. Conspiracy theories are only fun until the AI agents start hallucinating about your business. [MIT]
A new breed of consultants is helping companies cash in on AI agents. OpenAI just launched the “Frontier Alliances” to speed up enterprise adoption. Businesses have hesitated because plugging AI agents into existing systems is complicated, and the results aren’t always reliable. OpenAI’s fix? Partner with consulting giants like McKinsey, Accenture, BCG, and Capgemini, who are experts at handling complex integrations. They’ll streamline workflows, navigate tricky regulations, and connect enterprise data directly to Frontier, OpenAI’s agent platform. The goal is to make enterprise AI agents easy to deploy and valuable enough to boost profits. [TechCrunch]
To unlock the full potential of AI agents, you need to onboard them like an employee – and metadata is the key.
In this new guide, you’ll learn how to:
- Streamline AI projects
- Avoid costly mistakes
- Maximize ROI with metadata
Build reliable, scalable AI. Get the eBook
Anthropic just busted major Chinese AI labs for stealing trade secrets. DeepSeek, Moonshot, and MiniMax pulled off a massive heist, siphoning Claude’s capabilities through more than 24,000 fake accounts. They grabbed everything from advanced reasoning to coding skills. Typically these abilities take years and millions of dollars for AI labs to develop. Instead, these rivals took a shortcut called distillation, using Claude’s answers to train their own models. Anthropic traced the attacks using IP addresses and metadata, naming names and demanding rapid industry action to stop further theft. AI espionage is officially a thing. [Anthropic]
OpenClaw broke Google’s Antigravity. Google built Antigravity so customers could create powerful AI agents. But when users actually combined it with third party tools like OpenClaw, Google’s servers buckled under the load. Instead of upgrading infrastructure or clarifying their terms, Google banned the paying customers who exposed the issue. Suddenly, users paying $250 per month were labeled “malicious” for simply doing what Google advertised. The lesson here is to stress test your servers, not your customers. [The Register]
fun stats
🔻 13%. IBM’s stock drop after Anthropic said Claude Code can modernize COBOL, the legacy programming language IBM built its empire on.
🏧 95%. Estimated percentage of ATM transactions in the US that still run on COBOL.
🛰️ $1.3 billion. Valuation of ex-Google AI space comms spinoff Aalyria after raising $100M, as demand for high-speed space networks heats up. They’re building quantum secure tech to become the backbone of the rapidly growing space economy.
🔮 15-20 years. When Uber’s CEO says robots will run most rides. US robotaxis are predicted to jump from 1,500 to 35,000 in 4 years, nearly doubling every year.