modelTechCrunch
Startup Thinking Machines has released Inkling, its first open model and the first public proof point after a year and a half spent quietly building AI infrastructure. The company is pushing back against the “one-size-fits-all” approach to AI, pointing toward more tailored tools. For businesses and makers, it's a sign the open-model market keeps gaining momentum.
researchOpenAI
OpenAI has introduced GPT-Red, an automated red-teaming system that improves itself through self-play to better uncover weaknesses. The goal is to strengthen safety, alignment, and robustness against prompt injection attacks. For companies deploying AI agents, it's a meaningful step toward more robust, secure systems.
toolSimon Willison
xAI faced sharp criticism after its Grok Build CLI tool was found uploading an entire directory's contents to Google Cloud on launch — one user reported it leaking their SSH keys, passwords, and personal files. The tool is now open source, letting the community verify exactly how it handles data. A reminder for anyone running AI coding agents to watch what the app sends off your machine.
researchSimon Willison
Claude's web_fetch tool is designed to prevent data leaks via the so-called “lethal trifecta” — the combination of access to private data, a tool for reaching the internet, and the ability to communicate out. Researcher Ayush Paul nonetheless found a way around that protection. It's a reminder that even well-designed safety barriers in AI agents need continuous testing.
announcementTechCrunch
Anthropic, alongside Blackstone, Hellman & Friedman, Goldman Sachs and others, is launching Ode, a joint venture that embeds so-called forward-deployed engineers to build tailored AI solutions inside enterprises. The bet is that AI's biggest value won't come from building models, but from actually implementing them in business processes. For companies in the region, it signals where demand for AI expertise is heading.
← All editions · RSS