Darwinium adds AI agent fraud checks across journeys

Darwinium has introduced two fraud prevention features built around what it calls Intent Intelligence. The additions are aimed at transactions involving AI agents, as well as humans and bots.

The new tools, Journey Transition Probability and MCP Protection, are designed to assess behaviour across a digital journey rather than relying only on credentials or single events.

Darwinium is repositioning its platform around continuous assessment of what a user or software agent is trying to do at each step of an interaction. The approach is intended to help businesses detect suspicious changes before a login, account action or payment is completed.

Journey Transition Probability examines the order and timing of actions against a customer's usual behaviour, a company's wider traffic patterns and the type of journey under way. It is intended to expose activity that appears routine in isolation but unusual when viewed as part of the full sequence.

MCP Protection focuses on the external tools used by AI agents. MCP, or Model Context Protocol, connects AI agents to outside systems and services. Darwinium said its software links those tool calls back to the journey that led to them, so businesses can verify credentials, monitor actions and apply extra checks before higher-risk steps such as payments.

The move reflects a broader shift in online fraud and customer interactions as AI agents take on tasks previously carried out directly by users. Darwinium said authenticated customers can still be manipulated into scams, while authorised AI agents may begin with legitimate instructions but later behave in unexpected ways.

"An authenticated (human) customer can still be coached into sending money to a scammer. An authorized AI agent can start out doing exactly what a customer asked, then take an unexpected turn," said Michael Rodriguez, Chief Operating Officer, Darwinium.

He also outlined the company's wider platform strategy.

"We've spent years building a view of the full customer journey because those changes matter. Making Intent Intelligence technology central to our platform lets businesses assess what's happening as an interaction unfolds, whether a person or an agent is taking the next step," said Rodriguez.

Growing gap

Darwinium cited survey data showing the scale of concern among fraud and security teams. According to the company, 97% of fraud, risk and security leaders reported an increase in AI-driven attacks, while 36% said they believed they had effective fraud coverage across the full customer journey.

It also pointed to data from its own network suggesting agent-led activity remains difficult to identify. Only around one in four agentic transactions explicitly declares itself, while transactions involving agents are rejected nine times as often as other purchases, Darwinium said.

That creates a challenge for companies trying to distinguish legitimate automated activity from risky behaviour. As AI agents become more common in commerce and account management, fraud controls built around single checkpoints may struggle to judge whether a sequence of actions remains aligned with the original intent.

Darwinium said its existing Agent Intent Detection product already identifies AI agents, including those that do not declare themselves. The new additions expand that by combining activity from a customer's web or mobile session with an agent's tool calls into a single journey view.

Risk decisions are made through a business's existing content delivery network, the infrastructure used to handle website traffic, Darwinium said. That means the analysis is intended to sit within established web delivery systems rather than requiring a separate customer-facing process.

Customer view

Apollo.io, a Darwinium customer, said the shift reflects a change in how web traffic should be assessed as software agents become harder to distinguish from human users through standard declarations.

"Darwinium vastly improves our visibility into user intent and behavior across surfaces, whether the user is an agent or a human," said Jon Ferrari, Senior Manager, Fraud Prevention and Application Security, Apollo.io.

He said traditional indicators are becoming less useful as agent activity increases.

"Web traffic is at an inflection point where user-agent declarations and even statements of intent are becoming moot. We need to track whether behavior matches stated intent over time, including in aggregate, where activity spread across many requests can add up to an outcome that no individual request disclosed," said Ferrari.