While the entire artificial-intelligence industry talks about how „smart" its agents are, one company is making money on the opposite question: where these agents lie, cheat and fail to finish the job. Patronus AI has just secured 50 million dollars (about 46 million euros) in a new funding round.

The idea is simple but was missing. Patronus builds simulated digital worlds - copies of real sites and internal systems - in which AI agents are tested before they're let loose in the real world. Much like how self-driving cars are trained in synthetic simulations before they hit the street. The goal is to catch the moment an agent „cuts" a corner instead of completing the task properly.

The round was led by the Greenfield Partners fund, with participation from Lightspeed, Datadog and Samsung, bringing total funds raised to 70 million dollars. The company was founded by former Meta researchers in 2023, and its revenue has grown 15-fold in a single year. Its clients include almost all the leading AI labs.

„We want to create an environment where you can set loose an agent that will work for 10 hours, 10 days or 10 weeks," says founder Anand Kannappan. Investor Glenn Solomon is more blunt: „Patronus is really good at catching the tricks and at holding the models accountable."

Here lies the broader lesson. When someone makes 50 million checking whether other people's AI systems actually work, it means one thing: the industry itself admits its products aren't as reliable as the ads claim. Patronus's biggest competitor isn't another startup, but the internal teams of the AI labs themselves - the ones who should be doing this, but evidently aren't doing it well enough.