Let Anthropic take the wheel

Welcome to Runtime! Today: Anthropic's Claude AI model can now control a computer, and imagine the places it will go, Apple's Private Cloud Compute is open for testing, and the latest enterprise moves.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Ok computer

Right now enterprise AI is abuzz with the possibilities that agents could deliver for vendors that need a new source of growth and enterprises that would like to reduce costs. After all, something has to justify spending the last two years obsessing over generative AI technology.

Anthropic took agentic AI a step further this week, introducing a new capability in Claude 3.5 Sonnet that can take control of a computer and execute tasks. The company was quick to emphasize that this work is very preliminary and is far from ready for regular use, but it raises a number of possibilities and questions about how agentic AI could transform enterprise tech.

• "A vast amount of modern work happens via computers," Anthropic said in a blog post, and it's hard to argue with that.
• Anthropic believes Claude's new "computer use" capability is a breakthrough because it can apply its AI model to any piece of software, as opposed to working through a specific user interface, such as Claude's web tool or OpenAI's ChatGPT.
• "When a developer tasks Claude with using a piece of computer software and gives it the necessary access, Claude looks at screenshots of what’s visible to the user, then counts how many pixels vertically or horizontally it needs to move a cursor in order to click in the correct place," Anthropic said.
• Claude's score of 14.9% on computer-use tests provided by OSWorld is nowhere near as good as a human (most of us get between 70%-75%, according to Anthropic), "but it’s far higher than the 7.7% obtained by the next-best AI model in the same category."

Assuming Claude's computer-use feature becomes more reliable and easier to use (which is not a given), it could put a nail in the coffin of the RPA market and introduce all sorts of possibilities for other types of automated operations, some of which are good and some of which are bad.

• A number of companies working on agentic AI are hoping it can solve the limitations of RPA, or robotic-process automation.
• RPA requires users to constantly update the instructions they gave the system as new variables emerge, and computer-use AI wouldn't need that kind of maintenance.
• "Instead of hard-coded scripts, it actually understands what it’s looking at," wrote Paul Chada, co-founder of Doozer AI in a LinkedIn post spotted by CIO.
• Simon Willison also used Claude's new feature to write programs in C, query databases, and download updates to an open-source tool installed on his computer.

On the down side, computer-use AI could enable the spam bot or malware firehose of a criminal hacker's wildest dreams. And given how Claude scans screenshots of a computer's desktop to make decisions, it is quite vulnerable to prompt injections.

• "This easily automates the task of getting a machine to go to a website and download malware or provide secrets, which could scale attacks (more machines hacked in a shorter period of time)," wrote Rachel Tobac, CEO of Social Proof Security, on X.
• Security experts warned about AI supercharging these kinds of attacks during the early days of the generative AI boom, while emphasizing its benefits for defenders.
• Anthropic also acknowledged that malicious instructions could be coded into images on a website that when scanned by Claude could redirect it to perform a different task than the one it was originally given.
• "Those using the computer-use version of Claude in our public beta should take the relevant precautions to minimize these kinds of risks," it wrote, and provided developers with some guidelines.

"Our goal is for Claude to take pre-existing pieces of computer software and simply use them as a person would," Anthropic said.

• While there's nothing simple about what Anthropic has introduced, that goal does get at the heart of the debate over generative AI: how much work do we want to turn over to the machines?
• Enterprise tech already relies on a great deal of automation to keep sites running and businesses growing, and tools such as Claude could add another dimension to that strategy.

Bugs wanted

Apple piqued the interest of cloud and security researchers earlier this year when it introduced  Private Cloud Compute, and it is following through on a commitment to let them verify its claims. After letting a few trusted experts behind the curtain, Apple will now allow the public access to its architecture to put its claims to the test and find flaws, it announced Thursday.

"For the first time ever, we’ve created a Virtual Research Environment (VRE) for an Apple platform," Apple said. Apple's Private Cloud Compute was designed to offload compute-intensive AI tasks from iPhones and Macs to its cloud servers without exposing that data to the public internet, and it uses a soup-to-nuts assortment of Apple hardware and software to isolate those queries.

Apple will also award bounties to security researchers who find flaws in PCC, which is expected to go live next week. If you can find a flaw that would allow "arbitrary code execution with arbitrary entitlements" in PCC, Apple could pay you up to $1 million, which was a little more revenue than it made every 60 seconds in 2023.

Enterprise moves

Benjamin Revcolevschi is the new CEO of OVH, after serving as deputy CEO of the French cloud infrastructure provider for the last six months.

Amit Zavery is the new president, COO, and chief product officer of ServiceNow, following almost six years in prominent leadership roles at Google Cloud.

Scott Schools and Aaron Chatterji are the new chief compliance officer and chief economist, respectively, at OpenAI.

Justin Greenberger is the new chief customer officer at AuditBoard, joining the compliance software company after six years in a similar role at UiPath.

The Runtime roundup

A TSMC executive said Thursday that initial production yields at its new Arizona plant are slightly higher than comparable yields in Taiwan, according to Bloomberg, which is a huge vote of confidence in the chip foundry's ability to expand beyond its home country.

In more good news for TSMC, Nvidia CEO Jensen Huang said it has fixed the design flaw in its next-generation Blackwell GPU that led to manufacturing delays, and that "it was 100% Nvidia's fault."

IBM missed revenue expectations for the third quarter citing weak hardware and consulting revenue, and its stock fell more than 6% Thursday.

ServiceNow beat Wall Street's estimates for revenue and profit during the quarter, and its stock rose more than 5% Thursday.

Thanks for reading — see you Saturday!