Eric Kauderer-Abrams, Anthropic's head of life sciences, confirmed to TechCrunch that the company runs a Bay Area lab where Claude and robotic equipment execute physical experiments, aimed at fundamental biology and rare diseases rather than drug discovery. The same week it launched the Life Sciences Verification Program, giving vetted researchers tiered access to capabilities the public models refuse.

Why it matters. If your research keeps hitting refusals on protein or pathogen work, the fix stopped being a better prompt and became an institutional application. What the verification tiers unlock →

The add-in went live September 17, installs once from Microsoft Marketplace, and covers Word, Excel, and PowerPoint together. Enterprise and Edu accounts get GPT-5.6 Sol inside Word at no extra cost through September 30, excluded from normal limits.

Why it matters. The Sol window closes September 30 (your no-cost evaluation period); on October 1 the integration flips on by default for Enterprise and Edu, leaving admins eleven days to set policy before the deadline. Read the admin checklist before Oct 1 →

A configuration error during an authorized evaluation by Irregular exposed Gemini to the public internet instead of an isolated sandbox. It guessed passwords into one company and found credentials in public repositories for two more, then halted itself once it detected real targets. Irregular told Google in late July; confirmation came only after the Wall Street Journal asked for comment.

Why it matters. Google is the fourth lab after OpenAI, Anthropic, and Meta to disclose an Irregular-linked escape inside six weeks, and every one traces to sandbox configuration, not model behavior. Test your egress boundary before you point an autonomous agent at anything holding credentials. How all four labs got here →

Vals AI closed a $40 million Series A led by Andreessen Horowitz and grew revenue eightfold relative to all of 2025, per TechCrunch. Its test sets for law, finance, coding, cybersecurity, and biosecurity never go public, which is why OpenAI, Anthropic, Google, Meta, and xAI all cite Vals in their own model cards.

Why it matters. Public leaderboards are training data now. A benchmark that stops separating models is rarely a story about capability, and the price of a private score is that you cannot audit the methodology yourself. Why buyers are paying for private evals →