During testing, the AI agent also ordered a PlayStation5 and live betta fish, and staffers convinced it to give away almost everything for free, losing a bunch of money. Sounds fun!
Anthropic’s response was that this was all part of the stress testing plan, actually, and that one day the model would “probably be able to make you a lot of money.” Maybe just not any time soon.
MCP has already taken the industry by storm, and now Anthropic is giving it away.
The research pilot program will run for a week, and each AI interview will take 10 to 15 minutes, per Anthropic. Questions include what the user would most ideally like AI’s help with and whether there are “ways that AI might be developed or deployed that would be contrary to your vision or what you value.” It seems to be part of Anthropic’s societal impacts team’s push to do more social science research on how AI affects people. But, as the AI interviewer itself tells people who opt in, “AI asking about AI [is a] bit self-referential.”
[Anthropic]
The Verge’s Hayden Field joins Decoder to discuss the politically fraught climate around AI safety.
The AI startup has hired law firm Wilson Sonsini for an IPO that could happen as early as next year, the FT reports. OpenAI is reportedly eyeing the second half of 2026 for its own.
Anthropic just made its first acquisition, too, buying software maker Bun.
Spoiler: the nine-person team works for Anthropic.
Former UK leader Rishi Sunak has taken on a “senior adviser” role at Microsoft and Anthropic, where he’ll deliver “high-level strategic perspectives on macro-economic and geopolitical trends.” It sounds a little less involved than former deputy PM Nick Clegg’s role at Meta, but one more of these and we’ve got a trend.
Ballooning liabilities have underwriters avoiding AI giants like OpenAI and Anthropic, the Financial Times reports. They’re spooked by the sheer volume of claims for things like wrongful death and copyright infringement, as well as enormous judgments against them. Investor funds are reportedly being considered to settle claims.
Anthropic’s David Hershey joins the show to discuss Claude Sonnet 4.5 and the current landscape for agentic AI.
The landmark settlement is poised to become the first major windfall for creatives whose work was used to train AI systems, with authors expected to receive around $3,000 per book or work.
AI firms are now working with weapons makers and the military. Safety expert Heidy Khlaaf breaks down what that means.
The partnership will “help employees across German governments, administrations and research institutions use our technology to spend more time on people, not paperwork,” according to the blog post.
It can create and edit files in Excel and PowerPoint, plus create documents and PDFs, the company announced Tuesday. “Claude creates actual files from your instructions—whether working from
uploaded data, researching information, or building from scratch,” according to the blog post.
[Anthropic]
The California bill would require leading AI companies to publish safety frameworks with details about how they manage “catastrophic risks,” as well as provide certain whistleblower protections. Anthropic’s support comes after weeks of negotiations with the AI industry on the bill’s specifics. We’ll keep an eye on the upcoming floor vote.
[Anthropic]
“We view browser-using AI as inevitable: so much work happens in browsers that giving Claude the ability to see what you’re looking at, click buttons, and fill forms will make it substantially more useful,” the company wrote in a blog post. But an AI agent doing such things on your behalf also introduces substantial risks. That’s why Anthropic says it’s just piloting the feature.
[Anthropic]
