Solutions / Agent Risk Intelligence
Trusted-agent protocols confirm an AI agent is valid and operating within a user mandate. UMCA confirms the transaction it is about to make is still safe, catching prompt injection, social engineering through the agent, mandates obtained by deception, and velocity anomalies that mandate-level limits miss.
Every transaction arrives with agent-level context: agent ID, mandate hash or token, declared intent, and session identifiers. The platform ingests this alongside standard payment fields.
Agent reputation, historical outcomes, velocity across the agent's user base, counterparty history, and intent-vs-cart alignment are added to the risk picture.
Rules, ML models, and behavioural analytics evaluate the transaction against both standard fraud typologies and agent-specific typologies: prompt injection, social engineering, mule payouts within limits, mandate drift.
Return approve, review, or decline in under 30 milliseconds with a clear explanation of which agent-level signals contributed, suitable for policy, dispute, and regulator review.
For a deep dive, see Fraud Typologies in the Agentic Era.
Agent ID, mandate hash, declared intent, session ID, and agent reputation are first-class risk features. The platform models the agent as a distinct actor, not just metadata on a payment.
The declared intent at the start of an agent session is compared against each outgoing payment. Divergence, scope creep, and sudden recipient changes within a session are flagged as risk signals, which is how prompt injection usually manifests.
Reputation is built from historical outcomes across all users of that agent, not per customer. Velocity is tracked at agent, user, and session levels, so a new agent paying its first mule-merchant looks very different from a long-trusted agent.
Expose a consent and policy layer to your customers so they can define which agents are allowed, with what limits, and when step-up is required. The bank remains the owner of consent, UMCA provides the backend.
Every decision carries a human-readable explanation of which agent-level signals contributed and in what proportion. Suitable for customer dispute, policy review, and regulator inspection.
Compatible with major trusted-agent protocols and emerging agentic payment standards. Equally capable where no formal agent protocol exists yet, which covers most account-to-account rails and open banking.
Trusted-agent protocols answer two questions: is this agent valid, and does it have a mandate? They do not answer the third: is the transaction inside the mandate actually fraudulent? That is where Agent Risk Intelligence lives.
Full explanation: The Trust Gap: Why Agent Protocols Do Not Stop Fraud.
Card networks are deploying agent protocols on their rails. Account-to-account rails like Interac, SEPA Instant, FedNow, and open banking under PSD3 have no equivalent. UMCA provides agent-aware fraud scoring on all of them.
Read Agent Risk on A2A Rails.
AI agents are being granted the authority to pay on behalf of users across banking, e-commerce, and B2B. The surface area for fraud is new, and risk teams need a dedicated detection layer.
See the full guide: Agentic Fraud Prevention: A Guide for Banks and Payment Providers.
