In short
- The newly proposed “Agentic Risk Standard” separates AI services into automatic payment services protected by escrow and cash management services that require registration.
- In tests, writing reduces the loss of users by up to 61%, even if zero payments leave the writers unfulfilled.
- Accurate estimation of the failure rate remains a major challenge because over- and under-estimation leads to systemic risk.
Like AI assistants start managing payments, financial transactions, etc actionsThere is a growing concern about the financial risks that fall on the person behind the agent when their actions fail. A consortium of researchers say that the current methods of protecting AI do not eliminate the risk, and new ways of insurance should be considered.
Recently paperresearchers from Microsoft, Google DeepMind, Columbia University, and startups Virtuals Protocol and t54.ai developed the Agentic Risk Standard, a standard system designed to compensate users when an AI agent violates a service, fails to provide services, or loses money.
“Technical security can provide reliable assurance, while financial users often want concrete guarantees of results,” the paper said.
The authors argue that recent AI research focuses on improving the modeling process, including reduction favoritismcreating complex systems driveand make their choices easier to understand.
“These threats are by design and cannot be solved by security alone because the behavior of the agent is unstable,” he wrote. “To overcome this gap between the reliability of the model and the reliability of the users, we are proposing a framework for the management of risks.”
The Agentic Risk Standard adds financial security to how AI services are managed. For simple jobs where the user just pays for the job, the payment is held in escrow and released only after the job is confirmed. For high-risk activities that require an upfront cash flow, such as trading or currency exchange, the system brings the underwriter down. The underwriter assesses the risk, requires the service provider to post collateral, and reimburses the user if a hidden failure occurs.
The paper stated that non-pecuniary damages such as illusion, defamation, or emotional harm do not exist.
The researchers said that the system was tested using a prototype that tested 5,000 tests, adding that the test was small and not designed to show actual failure.
“These results encourage future work on the development of different types of failure risk models, comprehensive measurement of long-term failure under deployment-like conditions, and the design of labeling and bonding materials that remain robust under detector errors and best practices,” the study said.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
