Hacken & Ant Group Launch Guardrails — Preventing AI Agents from Making Unauthorized Transfers

Imagine an AI agent autonomously opening a wallet to send tokens or transferring money through a bank account. This isn't just a hypothetical scenario—it's technology already becoming reality. But how can we stop someone from tricking an agent into sending money to a hacker's account via a subtle prompt injection attack?

Until now, we have largely relied on static security audits to scan for code vulnerabilities before service deployment. However, to effectively control the unpredictable behavior of AI agents that change from moment to moment, we need a new type of execution guardrail—one that validates and intercepts policies in real-time, right before an action is executed.

Agents Compromised by a Single Prompt: Breaking the '5-Stage Attack Chain'

Services that delegate on-chain wallet permissions to autonomous AI agents are increasing rapidly. Yet, this convenience hides a critical threat: a single, cleverly manipulated prompt injection attack can cause an agent to lose complete control and transfer assets to a hacker's address.

Web3 security firm Hacken has precisely analyzed the process of an AI agent being manipulated by a hacker into unauthorized transfers as a 5-stage attack chain. It starts with the Injection phase, where the agent reads external documents containing hacking commands. This is followed by Hijacking, where the agent's tool permissions are stolen; Abuse, where the system is manipulated; Transaction Creation, where a fraudulent transaction record is crafted; and finally, Broadcast, where the malicious transaction is sent to the blockchain network.

To break this dangerous flow, Hacken has introduced the Transaction Policy Gateway. It serves as a real-time guardrail that intercepts and verifies transfer requests at the final junction before the agent sends the transaction to the blockchain. Even if the inside of an agent is compromised by a hacker, if it violates pre-defined safety rules at the final transmission step, the transaction is immediately blocked, effectively preventing asset leakage at the source.

Forget Compliance Prompts: Validate Real-Time Behavioral Paths

Simply tuning system prompts like "be a good agent" cannot perfectly prevent unexpected AI behavior. Agents constantly adjust their actions along unpredictable paths depending on the situation. This is precisely why academia and regulatory bodies are racing to adopt real-time guardrails that monitor agent activity as it happens.

The paper 'Policies on Paths,' released in March 2026, provides the theoretical foundation for this real-time verification. Researchers point out the need for an engine that analyzes the 'execution path' of actions the agent has taken to calculate the probability of policy violations in real-time. This means going beyond static, pre-set rules to track the real-time context of the workflow in which the agent is performing tasks.

The financial sector is also rapidly adopting this principle. Most notably, the Monetary Authority of Singapore (MAS) officially released the SAFR v1.0 guidelines in July 2026, which provide real-time protections for AI financial agents. To prevent unauthorized automatic asset transfer incidents in financial transaction loops, these guidelines mandate strong security rules for securing the execution path of agents in real-time.

Open-Source Real-Time Shields: SingGuard-NSFA and IRSB

Open-source solutions that developers can apply directly to their services are emerging as well. A prime example is the SingGuard-NSFA framework released by Ant Group. This tool inspects risks in a sandbox environment right before an agent takes action—similar to robotics or real-time trading—effectively blocking system shocks or API errors at the source.

Another notable project is IRSB, an open-source guardrail on GitHub. IRSB leverages EIP-7702, a new Ethereum standard, and cryptographic receipts to safely control an agent's financial transaction authority. Instead of handing over an entire private key to the agent, it establishes an on-chain defensive wall that allows funds to be executed safely only within defined limits.

Thanks to this, developers no longer need to build complex security logic from scratch. It has become easy to integrate a robust, real-time shield into agent systems that simultaneously prevents sudden malfunctions and external hacking threats.

Moving Beyond Static Audits to an Era of Real-Time Behavioral Shields

While security has previously focused on scanning for vulnerabilities before code deployment, it is now shifting toward a battle of guardrails that intercept and verify agent actions in real-time just before they occur. This is because no matter how precisely you craft prompts, you cannot perfectly control every unexpected action an agent might take.

Developers and builders looking to adopt AI agents in business must not rely solely on prompt tuning. You must first consider how to integrate a physical execution control framework that can block system malfunctions and financial threats in real-time. Creating a smart, real-time shield that tracks and controls execution paths to the end is the most certain key to building safe agent services.