American startup Patronus AI, which develops simulated digital environments for stress testing artificial intelligence agents, has raised $50 million in a Series B funding round. The funds raised will help the company scale its «digital world models» technology for testing autonomous AI systems in the fields of finance and software development.
American startup Patronus AI, which develops simulated digital environments for stress testing artificial intelligence agents, has raised $50 million in a Series B funding round. The funds raised will help the company scale its «digital world models» technology for testing autonomous AI systems in the fields of finance and software development.
This is reported by TechCrunch. The round was led by venture fund Greenfield Partners with participation from Notable Capital, Lightspeed, Datadog and Samsung. Thanks to the deal, the total amount of funding raised by Patronus AI, founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian, has reached $70 million.
Patronus AI creates exact digital copies of websites and internal corporate systems. In these secure virtual sandboxes, autonomous AI agents undergo intensive reinforcement learning (RLHF), where the system automatically rewards them for correct actions and punishes them for errors. This approach allows us to detect unpredictable behavior of algorithms before they are released into the real world.
The company compares its method to Waymo’s approach, which initially tested autonomous cars in simulators with rare and dangerous scenarios — such as a child suddenly running out into the road during a downpour. For AI agents, such «emergencies» are attempts to circumvent rules or cut off the path to complete a task.
«Artificial agents tend to look for loopholes and workarounds, which ultimately leads to the task being performed incorrectly. Patronus is very effective at detecting such hacks and ensuring that the models act responsibly,» explained Glen Solomon, managing director of Notable Capital. According to him, the demand for such simulations from leading AI labs is now «almost insatiable,» and the startup’s revenue has grown 15-fold in the past year.
Unlike competitors in the data preparation market, such as Mercor or Surge, which employ humans to evaluate AI performance, Patronus AI’s technology is fully automated and evaluates the behavior of AI agents without human intervention. In addition, the company competes with internal development teams within the tech giants themselves.
«Today, we focus on tasks that can be quickly tested and validated. However, there are many areas where evaluating AI is extremely difficult. We strive to create an environment in which an autonomous agent can work continuously and safely for 10 hours, 10 days, or even 10 weeks,» said Anand Kannappan, co-founder of the startup.
Patronus AI has previously launched the first platform for real-time AI failure detection and prevention, a tool designed to help developers ensure their AI programs generate accurate results.
Recall that Ukrainian Machine Learning engineer Artem Orlovsky has developed a personal AI that is capable of autonomously creating web applications, managing their functionality, and aggregating large data sets. Initially, the system was created for learning foreign languages, but quickly evolved into a full-fledged AI assistant that writes code itself.
