AI Agent4 months ago
News Editor From Blindly Saying Yes to Verifiable Approval: How Sigil Adds a Security Guardrail to AI Agents
As AI agents move from answering questions to taking real actions on behalf of users, authorization is becoming a critical security layer. In this article, imToken’s new “Sign” narrative is examined through Sigil, an early proof-of-concept product designed to sit between AI agents and wallets. The core issue is that many agents still rely on vague chat-based approvals such as “Yes” or “Confirm,” even when the underlying action may involve fund transfers, token approvals, contract calls, code execution, or access to sensitive services. That model creates three major risks: users often cannot see the exact parameters of what they are approving, a chat reply is not equivalent to a cryptographic signature by the account owner, and the approval interface itself can be manipulated by the same agent that initiates the request. Sigil’s answer is to rebind display, signing, and execution into a verifiable flow. Low-risk actions can proceed under predefined policies, while sensitive operations trigger a pause and generate a structured approval card sent through Telegram. The user must then confirm with Passkey and biometric authentication before the action can continue. The article also highlights Sigil’s three protection layers: clear presentation of real request parameters, proof that only the actual user can approve, and an isolated confirmation module anchored on-chain and rendered in a sandbox to prevent interface forgery. While the need is most obvious in crypto, the framework may extend to email, files, terminals, identity usage, payments, and broader agent permissions.