Virtuals Protocol reposted this
⏰ 😁 Excited to share our new paper: Quantifying Trust: Financial Risk Management for Trustworthy AI Agents, together with Tianyi Peng Chi Wang t54 Labs Virtuals Protocol 📌 paper: https://lnkd.in/eXZDiqfm 📌 github: https://lnkd.in/ey2aNsrv 📌 website: http://t54.ai/ars 📌 featured in Fortune https://lnkd.in/e5S4rQUQ We propose the Agentic Risk Standard (ARS), a settlement-layer protocol that uses escrow, underwriting, and collateral to provide enforceable guarantees for users when AI agents transact on their behalf. Most trustworthy AI research focuses on model-internal properties: alignment, robustness, interpretability, bias mitigation. These have produced real progress, but they share a common ceiling. Large language models are stochastic. No amount of alignment training reduces failure probability to zero. For a chatbot, that's a tolerable property. For an agent executing a trade, filing a tax return, or moving money through a financial API, a single failure can cause harm that vastly exceeds the value of the task itself. We call this the guarantee gap. Model safety techniques offer probabilistic reliability. Users in high-stakes settings need enforceable guarantees over outcomes. Better training shrinks the gap. It cannot close it. Our approach: ➡️ stop trying to eliminate failure, and instead specify how its financial consequences should be handled ⬅️ . This isn't a new question outside of AI. Construction uses performance bonds. Medicine uses malpractice insurance. Financial markets use margin requirements and clearinghouses. None of these mechanisms make the underlying activity safer in a technical sense. They make the downside contractible. ARS applies the same logic to agent transactions. Each job is anchored by a signed structured agreement and traverses a deterministic lifecycle, with explicit authorization predicates governing every fund-moving action. Fee-only jobs are handled with escrow. Fund-involving jobs add an underwriting layer that prices risk, requires provider collateral, and commits to reimbursement under explicit failure triggers. ⁉️ The harder problem isn't the protocol. It's risk modeling. ARS only works if an underwriter can estimate failure probabilities and loss magnitudes with enough accuracy to price coverage without becoming insolvent or pricing out legitimate users. For some failure modes (unauthorized transfers, trading misexecution) loss maps naturally to monetary terms. For others (hallucination, biased outputs, privacy harm) the harm is real but lacks an agreed monetary metric. Defining contractible proxies for these harms is a key open research direction.
