Anthropic Talks With Pentagon Stall Over Claude AI Guardrails

Описание: Anthropic PBC’s talks about extending a contract with the Pentagon are being held up over additional protections the artificial intelligence company wants to put on its Claude tool, a person familiar with the matter said. Anthropic wants to put guardrails in place to stop Claude from being used for mass surveillance of Americans or to develop weapons that can be deployed without a human involved, the person said, asking not to be identified because the negotiations are private. The Pentagon wants to be able to use Claude as long as its deployment doesn’t break the law. Bloomberg News Tech and National Security Reporter Katrina Manson joins Bloomberg Businessweek Daily to discuss. She speaks with Carol Massar and Tim Stenovec.  Anthropic PBC is releasing a new artificial intelligence model that’s intended to be better at using people’s computers in increasingly complicated ways, building on the startup’s efforts to make AI tools more effective at streamlining tasks. 

Claude Sonnet 4.6, set to be rolled out on Tuesday, can carry out actions on a computer that require multiple steps, such as filling out web forms and then coordinating information across several browser tabs, the company said. “The model certainly still lags behind the most skilled humans at using computers,” Anthropic said in a blog post. “But the rate of progress is remarkable nonetheless.”

Anthropic first introduced a “computer use” option in late 2024 that enabled an AI system to analyze users’ screens and take actions for them by browsing the web. Since then, competitors such as OpenAI and Alphabet Inc.’s Google have unveiled AI models to control a computer and complete tasks — particularly mundane ones traditionally carried out by humans — at a user’s command. OpenAI, in particular, recently hired the creator of OpenClaw, a popular open-source tool that runs on a user’s computer and, with the help of a connected AI model, can send emails or make restaurant reservations.

Sonnet 4.6 is now the default option for those who use its Claude chatbot for free or pay for a Pro subscription plan. Anthropic said the new model is also more reliable than its predecessor at coding, which has long been a key area of focus for the company. 

The Claude maker’s efforts to expand beyond its success with software developers have rattled Wall Street in recent weeks. Anthropic’s quiet release of a tool to automate certain legal work helped spark a market meltdown earlier this month, particularly among software companies that investors fear may eventually be rendered obsolete. Shares of financial services also slumped after Anthropic released a new version of its Opus model that’s meant to be better at financial research. The reactions reflect broader concerns about which companies and services will eventually be disrupted by AI.

Though designing AI models to perform a greater variety of actions could make them much more valuable to users, this approach comes with a new set of risks, too. Ceding control to such software can make people vulnerable to security incidents such as prompt injection attacks, where an AI model is manipulated by a malicious command. Anthropic said Sonnet 4.6 is much better than the AI model that preceded it,
--------
Watch Bloomberg Radio LIVE on YouTube
Weekdays 7am-6pm ET
WATCH HERE: http://bit.ly/3vTiACF

Follow us on X:   / bloombergradio  

Subscribe to our Podcasts:

Bloomberg Daybreak: http://bit.ly/3DWYoAN
Bloomberg Surveillance: http://bit.ly/3OPtReI
Bloomberg Intelligence: http://bit.ly/3YrBfOi
Balance of Power: http://bit.ly/3OO8eLC
Bloomberg Businessweek: http://bit.ly/3IPl60i

Listen on Apple CarPlay and Android Auto with the Bloomberg Business app:
Apple CarPlay: https://apple.co/486mghI
Android Auto: https://bit.ly/49benZy

Visit our YouTube channels:
Bloomberg Podcasts:    / bloombergpodcasts  
Bloomberg Television:    / @markets  
Bloomberg Originals:    / bloomberg  
Quicktake:    / @bloombergquicktake