Darwinium introduces two intent intelligence features to detect fraud by AI agents

Fraud prevention startup Darwinium UK Ltd. today launched two new features that make intent intelligence the foundation of its platform.

Journey Transition Probability characterizes unusual changes in the steps a person, bot, or artificial intelligence agent takes on the path to a sign-in, account change, or payment. MCP Protection monitors the tools agents use to complete tasks.

The release targets a problem made more difficult by agent trading, with only about one in four agent transactions on Darwinium’s network self-declaring. Purchases made through brokers are nine times more likely to be rejected than others.

In a Darwinium survey of 500 leading fraud, risk and security companies in the US and UK, 97% reported an increase in AI-driven attacks. Only 36% believed they had effective fraud coverage throughout the customer journey. The company argues that companies need a way to detect legitimate agent activity without giving rise to risky actions.

An authorized AI agent “can initially do exactly what a customer asks and then take an unexpected turn,” said Michael Rodriguez, chief operating officer at Darwinium. He also expressed the same thing about authenticated human customers who can still be tricked into sending money to a scammer. That’s why Darwinium has spent years mapping the entire customer journey, Rodriguez said. By focusing on intent intelligence, companies can evaluate an interaction as it unfolds.

The company launched its agent intent tools in March, and its existing agent intent detection product is already identifying AI agents even if they don’t announce themselves. With today’s update, the tool calls an agent makes using the Model Context Protocol are integrated into the same journey as a customer’s web and mobile activities. A model evaluates each step. Risk decisions run through the content delivery network that the company already uses for its website traffic.

Within this combined view, journey transition probability measures the order and timing of actions based on the customer’s typical behavior. The company’s own traffic and the type of journey in progress also play a role. A step that looks normal on its own can seem suspicious when you consider the entire journey.

MCP Protection takes care of the agent side by linking each tool call to the journey that led to it. This allows companies to verify an agent’s credentials and monitor what the agent is doing once they are working. A payment or other riskier action may be withheld for additional review.

Jon Ferrari, senior manager of fraud prevention and application security at Darwinium client Apollo.io, said web traffic has “reached a tipping point where user-agent declarations and even letters of intent are becoming obsolete.” His team needs to track over time whether the behavior is consistent with the stated intent, he said. This also includes aggregate activity, where many requests can lead to an outcome where no individual request is disclosed.

Darwinium is a venture capital-backed startup whose investors include US Venture Partners, Blackbird Ventures Pty. Ltd., Airtree Ventures Pty. Ltd. and Accomplice include. The company announced an $18 million round in October 2023.

Image: Darwinium