On an episode of The Verge’s “Decoder” podcast, Suleyman said that in Anthropic’s Claude Constitution, they “speculate about its consciousness and whether it has those feelings and is aware.” He called it both “dangerous” and a “philosophical failing,” adding, “We want AIs to be controllable, contained, accountable, aligned tools that serve humanity.”
RSI is also defined as an “AI system capable of fully autonomously designing and developing its own successor,” per Anthropic’s blog post. “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”
[Anthropic]
With this expansion of Anthropic’s Project Glasswing initiative, organizations in “several industries that weren’t well represented” in the initial cohort, like power, water, and healthcare, will get access to the model so they can use it to find security vulnerabilities.
[Anthropic]
The $65 billion Series H round gives Anthropic an eye-watering $900 billion valuation. That gives the company a higher valuation than OpenAI’s last valuation of $730 billion, according to The New York Times.
Anthropic says the funds will go toward advancing safety research, expanding compute, and scaling its products.
Anthropic’s fight with the Pentagon highlights the risks of autonomous warfare — but obscures just how close it is.
Upon request, “qualifying” customers can use things like skills, a Claude harness, and a threat model builder, Anthropic says as part of a bigger update about Project Glasswing.
Anthropic also plans to expand Project Glassing to “additional partners” and has published a dashboard of open source vulnerabilities disclosed by Mythos Preview.
[Anthropic]
Apparently, that SpaceX $15 billion per year megadeal isn’t even enough capacity for Claude, as The Information reports Anthropic is in early talks to rent Azure servers with Microsoft’s chips, and that “Anthropic has been steadily increasing its Azure usage.”
Like OpenAI, Microsoft’s arrangement with Anthropic runs hot and cold, but its Maia 200 chips are designed to help run existing models like Claude, even if they aren’t as fast at helping to train new ones.
[The Information]
Karpathy, who had also been on the founding team of OpenAI, says he will be working on R&D at Anthropic. Previously, he had been working on “new kind of school that is AI native,” and he says he’s still “deeply passionate about education” and plans to go back to it “in time.”
Researchers at the security firm Calif say they used Anthropic’s cybersecurity AI to create a privilege escalation exploit, the Wall Street Journal reports:
Last September, Apple said it leveraged its hardware and operating system expertise into a technology called Memory Integrity Enforcement (MIE), which it described as “the culmination of an unprecedented design and engineering effort, spanning half a decade.” With Claude, building the code that exploited the two MacOS bugs took five days, Calif says.
On Wednesday, the AISI, which evaluates AI models for the British government, said both Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 showed progress well above previous trends on cybersecurity testing. Separately, XBOW released data suggesting “frontier models have taken a major step forward in vulnerability discovery.”
Meanwhile, Microsoft said its multi-model agentic setup, MDASH, was used to discover 16 CVEs in this week’s Patch Tuesday updates and is the leader on the CyberGym security evaluation framework.
It’s a package of “connectors,” installed via a toggle switch, that allows Claude to work inside tools like Intuit Quickbooks, PayPal, Docusign, HugSpot, Google Workspace, and Microsoft 365. “It can plan payroll, close the month, run a sales campaign, chase invoices, and more,” per Anthropic’s blog post.
[Anthropic]
As the AI adoption in the legal field ramps up, Anthropic has announced that its AI chatbot can now connect to many of the apps used by lawyers, including DocuSign, Box, Thomson Reuters, Harvey, and more. “Claude reviews contracts, surfaces case law, and drafts across the tools your team already uses,” according to Anthropic.
The AI startup says “dreaming” will allow Claude to review previous sessions to “find patterns and help agents self-improve.” This feature, which is rolling out in research preview, is supposed to help AI agents identify frequent mistakes, spot tasks they might converge on, and understand a team’s preferences.
Anthropic is doubling five-hour rate limits for many Claude Code users, removing Claude Code’s peak hours limit reduction, and significantly increasing API rate limits for Claude Opus models, starting today.
It credits the capacity to a new deal with SpaceX “to use all of the compute capacity at their Colossus 1 data center” in Memphis, noting recent announcements with Amazon, Google, and Microsoft.
Asked by Senator Jacky Rosen (D-Nevada) in a Senate Armed Services Committee hearing about the DoD’s dispute with Anthropic and whether he could guarantee a human would be in the loop on any targeting decisions made with AI, Hegseth focused on Amodei and his company’s refusal to “accept our terms of service.”
Claude Security uses the Opus 4.7 model to scan a business’s codebase for vulnerabilities and issue a fix. This tool is rolling out to enterprise customers globally and isn’t to be confused with Anthropic’s Mythos, a powerful AI model that can identify and exploit vulnerabilities across operating systems and web browsers.
Initially, Google will invest $10 billion, but could pour up to $30 billion more into Anthropic if it meets certain performance targets, according to Bloomberg.
Amazon, which had already invested $8 billion in Anthropic before this week, also announced new investments into the company. It invested $5 billion on Monday and could commit “up to an additional $20 billion in the future.”
