Darwinium Launches Intent Tools to Detect Fraud by AI Agents
New Journey Transition Probability and MCP Protection features analyze agent steps and tool calls to flag risky behavior in real time.
Key points
- Darwinium releases Journey Transition Probability and MCP Protection to track AI agent behavior.
- Only 25 percent of agentic transactions self-declare, while rejections are nine times higher.
- A survey found 97 percent of leaders see rising AI attacks but only 36 percent feel covered.
Fraud prevention startup Darwinium UK Ltd. introduced two new capabilities today that place intent intelligence at the core of its security platform. The company aims to address rising fraud risks associated with artificial intelligence agents in commerce.
The release includes Journey Transition Probability, which flags unusual changes in steps taken during logins or payments, and MCP Protection, which monitors the tools agents use to execute tasks. These features help businesses assess interactions as they unfold rather than after the fact.
Context and Background
Agentic commerce presents management challenges because only about one in four transactions on Darwinium’s network self-declares as being driven by an agent, according to SiliconANGLE. Purchases involving agents are rejected nine times as often as other transactions, indicating higher inherent risk in automated interactions.
Darwinium conducted a survey of 500 fraud, risk and security leaders in the U.S. and the U.K. The results showed that 97 percent reported an increase in AI-driven attacks. However, only 36 percent believed they had effective fraud coverage across the entire customer journey, highlighting a significant protection gap.
Who Is Affected
Michael Rodriguez, chief operating officer at Darwinium, noted that authorized AI agents can start with legitimate requests before taking unexpected turns. He drew a parallel to authenticated human customers who may be coached into sending money to scammers, emphasizing that intent can shift even after initial verification.
The company launched its initial agent intent tools in March, with an existing product that identifies AI agents even when they do not announce themselves. Today’s update integrates Model Context Protocol tool calls into the same journey view as web and mobile activity, allowing a single model to assess every step.
What should you do?
Risk decisions run through the content delivery network that businesses already use for website traffic, according to the report. Journey Transition Probability measures action order and timing against typical customer behavior, while MCP Protection links each tool call to the journey that led to it. This combined view allows ordinary steps to be flagged if the full path appears suspicious.
What to do and how to stay safe: Darwinium
- Review current fraud coverage to ensure it extends across the full customer journey, not just initial login.
- Monitor for unusual changes in step sequences during account changes or payments, even from authenticated users.
- Verify that security tools can identify AI agents that do not self-declare during transactions.
- Wait for vendor updates before implementing new detection rules for Model Context Protocol tool calls.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What are the two new capabilities launched by Darwinium?
Journey Transition Probability flags unusual step changes, and MCP Protection watches the tools agents use to carry out tasks.
How common are AI-driven attacks according to the survey?
Ninety-seven percent of surveyed fraud, risk and security leaders in the U.S. and U.K. reported an increase in AI-driven attacks.
Why is agentic commerce harder to manage for fraud teams?
Only about one in four agentic transactions self-declares, and purchases involving agents are rejected nine times as often as others.
