|
Join The Information for AI Agenda Live on Wednesday, September 23, at SFMOMA. We’re bringing together the industry’s most influential researchers, founders, and investors to preview the next wave of innovation—from breakthroughs in continual learning models to the latest bleeding-edge applications. Learn more here.
Welcome back!
Anthropic’s AI is pricey, but customers are finding they can use it more efficiently by adding software known as a harness to control the way AI models operate.
Visa in April developed a harness that reduces the cost of and time it takes for Anthropic’s Mythos AI to find and fix cybersecurity vulnerabilities in its codebases significantly, the company’s president of technology Rajat Taneja told me.
“Using the model through this harness, we have found, is more effective than using the model by itself,” he said.
The company declined to share specific figures about cost and time reductions but pointed to a July study that showed a similar harness could cut the cost of using AI models by as much as 60%.
A harness is software that hands the model tools, such as access to GitHub and Slack; feeds it context about the user and their preferences; executes its actions on your computer or in the cloud, and can loop it back around when it fails.
Instead of an employee telling Mythos what to do, the harness software Visa developed breaks up the bug-finding and fixing process into 11 stages and handles some of the work up front, such as mapping out which parts of a company’s codebases should be scanned. The harness also double-checks the model’s work in case it identifies vulnerabilities that are false positives.
A Visa spokesperson said developing the harness was “a significant undertaking” that took nearly 100 engineers more than three months.
Visa has open-sourced the software, the Visa Vulnerability Agentic Harness, so other companies can use it. And on Thursday, the company updated the software so that users can easily switch between Anthropic, OpenAI, and open weight models such as Kimi or use a combination of those models.
Taneja said the harness noticeably reduces the time between when the Mythos model finds a vulnerability and actually fixes it, a metric called “time to adapt.”
Companies such as Globant, Coinbase, Shopify, and Ramp have also made their own harnesses that can also switch between different AI models for coding work and other white collar tasks.
Taneja declined to share Visa’s AI budget but said the company uses trillions of tokens monthly. That means Visa is assuredly spending millions of dollars monthly—for example, using two trillion tokens of Anthropic’s Claude Opus 5 would cost about $7.7 million per month, according to The Information’s calculation using blended prices listed by Artificial Analysis. That spending is set to increase as Taneja sees internal AI usage “growing at a pace that is unlike anything I have seen because we are adopting it everywhere.”
Several months ago, Visa became part of Anthropic’s Project Glasswing, in which the AI firm gave about 40 businesses early access to its super-powerful Mythos AI model. Mythos has still not been released to the public, though a commercial version called Fable 5 is available.
Visa handles more than 1 billion financial transactions daily, Taneja said, “so security of this system, and cybersecurity, is job number one.”
|
