As AI agents increasingly make purchases, payments, and account changes on behalf of consumers, banks face a new verification gap: confirming that an agent is legitimate does not necessarily mean the customer has authorised the specific action it is taking. US-based fraud prevention company Incognia has introduced AI Agent Detection to address this gap by evaluating the agent and the action separately. Alongside the launch, the company is extending its web risk intelligence with Web Behavioral Biometrics and introducing an MCP Server for AI-assisted fraud investigations.
Incognia separates agent verification from action approval
AI Agent Detection combines signed-request verification, agent-origin classification, and bot and automation detection with Incognia's wider risk intelligence. Where an action requires additional assurance, institutions can check it against signals from the customer's trusted devices, account activity, device integrity, network, and location. André Ferraz, co-founder and CEO of Incognia, said that confirming a request came from a legitimate agent is not the same as knowing a particular action should be permitted. He added that businesses need to link each request back to the real customer and its surrounding context, while keeping their existing risk controls in place.
The company's web offering now adds behavioural signals such as mouse movement, keyboard interaction, and copy-and-paste activity to its existing browser, device, automation, and location data. When further checks are needed, a web session can be connected to a trusted mobile device to provide additional evidence about whether the person behind the interaction is genuine. Incognia combines apartment-level location intelligence with device and behavioural signals to recognise trusted users across sessions, and works with companies in mobility, marketplaces, and financial services.
Read-only MCP Server supports fraud investigations
Through the new MCP Server, fraud analysts can use compatible AI applications to retrieve evidence, identify links between devices and accounts, and organise findings for review by asking questions in natural language. The initial version is read-only, so enforcement decisions remain under human control. The launch comes as lenders adopt agentic AI without consistent safeguards. Jim Mortenson, Fraud Strategic Advisor at Datos Insights, said many firms lack reliable verification, detection, and governance frameworks for these risks. He added that investing in control infrastructure now could strengthen long-term resilience and reduce the likelihood of costly redesigns.
