Imagine an AI trading bot executes a bad trade, or an autonomous payments agent sends the wrong amount — who covers the loss? A coalition of researchers from Microsoft, Google DeepMind, Columbia University and startups Virtuals Protocol and t54.ai say current AI-safety work leaves that human tail risk exposed, and they’re proposing an insurance-style fix aimed at the settlement layer.
In a new paper, the team introduces the Agentic Risk Standard: a framework that layers financial safeguards onto how AI agents carry out transactions. Their core point is simple but important for any system that lets AI touch money: technical fixes can only make behavior probabilistically better, not guarantee outcomes. For high-stakes financial interactions—trading, currency exchange, or releasing funds up front—that gap matters.
How the Agentic Risk Standard works
- Low-risk jobs: If a task only involves paying a service fee, funds are held in escrow and released only after the user confirms the work.
- Higher-risk jobs: For tasks that expose users to upfront financial loss, an underwriter assesses risk. The underwriter can require the service provider to post collateral and commits to repaying the user if a covered failure occurs.
The authors contrast this approach with most current AI research, which concentrates on improving model behavior—bias reduction, robustness to manipulation, and interpretability. “These risks are fundamentally product-level and cannot be eliminated by technical safeguards alone because agent behavior is inherently stochastic,” they write. The Agentic Risk Standard is intended to complement model-level safety by providing user-facing assurances via risk management and financial backstops.
What the proposal does not cover
The paper explicitly excludes many non-financial harms—hallucinations that don’t produce monetary loss, defamation, or psychological harms—focusing instead on monetary loss and settlement mechanisms.
Proof-of-concept and limits
The team ran a simulation with 5,000 trials to test the idea, but they stress the experiment was limited and not representative of real-world failure rates. The paper calls for more work on realistic risk models, deployment-like failure measurements, and robust underwriting and collateral schemes that hold up to detector errors and strategic behavior.
Why this matters for crypto and DeFi
The proposal reads like a natural fit for systems where money, automated agents, and on-chain settlement intersect. Escrow, collateral posting and underwriting are familiar primitives in crypto and decentralized finance, and the Agentic Risk Standard maps those primitives to a liability and consumer-protection layer for agentic services. As AI agents increasingly handle payments and trades, frameworks like this could define how responsibility and financial guarantees are structured—whether on-chain, off-chain, or in hybrid systems.
Bottom line: as AI agents take on transactional roles, the authors argue the industry should pair model improvements with financial-risk engineering. Technical reliability is necessary but insufficient; users in high-stakes settings will also need enforceable, monetary guarantees when agents fumble.
Read more AI-generated news on: undefined/news
