A joint research report by Bitget and SlowMist looks at the risks that arise as artificial intelligence systems begin to execute transactions autonomously. As trading activity enters this “agentic” phase, where systems move beyond analysis and into action, a new category of risk arises—one that traditional security models were not designed to address. The report highlights a fundamental shift: Once AI moves from advisory roles to execution, bugs and exploits are no longer isolated events. They can trigger immediate and irreversible financial results. In crypto markets, where transactions are settled instantly, a compromised or misdirected agent can act faster than a human can intervene. «AI no longer just interprets the markets, but actively participates,» said Gracy Chen, CEO of Bitget. «That completely changes the nature of the risk. The question is no longer how smart these systems are, but how safely they are allowed to operate.» According to the research, agent-based systems introduce new attack surfaces at multiple levels – from data entered into models to execution paths. Prompts can influence decision-making, malicious plugins can change behavior, and APIs with too many permissions can expose capital to unintended actions. These risks are amplified by the always-on nature of autonomous agents, which operate continuously without direct user supervision. Rather than treating these issues as isolated vulnerabilities, the report frames them as systemic. Security in the agentic age must go beyond the application level and extend to the architecture of how AI systems interact with capital. Bitget's approach reflects this change. The platform separates asset intelligence, execution and authorization into distinct layers, reducing the likelihood of a single point of failure triggering unintended transactions. Permissions are structured according to the principle of least access required, and transaction simulation and verification processes are introduced before the execution is completed. These controls are designed to ensure that even when AI agents operate autonomously, their scope remains defined and limited. SlowMist's analysis highlights the need for a closed-loop security model where risks are managed before, during and after execution. Continuous monitoring, limited permissions, and verifiable transaction flows form the basis of this framework, transforming security from a reactive process to an integrated element of system design. The findings point to a broader reality where AI agents are becoming more integrated into trading, asset management and onchain activities; thus, the boundary between user intent and system execution becomes increasingly abstract. In such an environment, reliability is no longer solely determined by performance, but by how well systems can operate within controlled limits. As financial activity becomes more automated and interconnected, infrastructure must be designed not just for speed and access, but also for containment and resilience. The joint report serves as a reference point for platforms, developers and users navigating this transition, highlighting that the next stage of financial innovation will depend as much on secure execution as it does on systems intelligence. The full report is here. Disclaimer: This article is provided for general informational purposes only and does not constitute investment, legal or financial advice, nor an offer or solicitation to buy or sell financial instruments or digital assets. Any opinions expressed are based on current market observations and are subject to change. Past performance is not indicative of future results. Digital assets are volatile and may not be suitable for all investors. Readers should conduct their own independent research and seek professional advice before making any investment decisions. Restrictions may apply. This content is intended for global users. Bitget may restrict or limit users' access to its services. This material is provided for informational purposes only and does not constitute financial advice. REVIEW Terms of Use Bitget.