teissTalk: Evolving identity and fraud models beyond human-only assumptions

Views on news

Mastercard has launched Agent Pay for Machines (AP4M), a new service that will allow these transactions to be permissioned, orchestrated and settled at machine speed across its global payments network. “Machine payments can make it possible for services to be bought and sold among agents at fundamentally different scales than payments today – very high volumes, very small values, very fast and at extremely low latency,” said Jorn Lambert, Mastercard’s chief product officer. AI agents are no longer just assisting decisions. They are able to act on human intent, coordinate services and complete transactions that are bespoke for their users. The challenge will be to validate at machine speed that the agent is truly making the payment on behalf of the human who owns the account. To achieve that, authorisations and identity checks must continuously work in the background. However, before these models become commercially available, there should be more guardrails put in place to make the technology safe and reliable for users.  We are reaching a stage where cyber security fundamentals must be applied not just to humans but agents as well. And the responsibility will lie with the service provider to ensure that what the agents are doing is in line with the intent of the human user. While MCPs are key to agentic AI, a survey earlier this year found that 40 per cent of publicly available MCP servers didn’t force authentication. However, finance, as a highly regulated industry, has a stronger focus on security and guardrails than most other sectors. 

Identity management is key to the security of agentic AI

Small language models or domain specific models are more secure than LLMs thanks to intent modelling and mapping, prompt injection protection, governance and fine-grained authorisation capabilities, and they often come with small orchestration models that can enforce delegation properly too. Another important aspect of agentic systems is how capabilities are put in front of MCPs to ensure transactions are discovered and brokered correctly.  Standards are being developed now by IETF and NIST to manage agentic intent. But agent behaviour is harder to monitor and some behaviour data used with humans can become irrelevant – agents can act as humans and, for example, tick zebra crossings in pictures to prove they are a human user. Unique identity management can address agent-specific security issues but it may also lead to the emergence of a black market of unique identities. Agents can’t be given human identity, but non-human identity is not possible either thanks to their non-deterministic nature. However, auth has already been extended to AI agents, giving them distinct digital identities, identifying the user and the agent, the authority assigned to the agent and its delegated scope. There are also security brokers embedded in these systems before the MCP. In Ping Identity, for example, these controls are already in place. DPIA (Data Protection Impact Assessment) can serve as yet another layer of agentic AI security. 

The panel’s advice

• Zero trust for agents, verified trust for humans.  
• While previous controls such as user behaviour remain relevant for agentic AI too, detection must adjust to the new technology.
• Agentic AI risks, being more rule-based, may be easier to manage than human-related ones.
• With EUDI coming out soon, it’ll be possible to link verifiable credentials to agents.
• Find the threat modelling – whether open source or commercial – that suits your business.
• Three important steps for agentic AI deployments:  learn more about AI agents by looking at Gartner’s six6 agent types; work hard on discovery; and focus on intent and purpose.
• To learn more about MCPs, click here.