CertiK: Agentic AI pivots from security assistant to autonomous operator

The transition marks a departure from traditional machine learning, which largely flagged anomalies for human review. According to a new report from CertiK, modern agentic systems can reason through multi-step threats, interface with external APIs, and execute responses in live environments. In a security operations center, this means an agent could theoretically identify a compromised account, analyze device history, and initiate a suspension—recording its own logic for later verification.

This shift is largely driven by the extreme velocity of Web3 exploits. With flash loan attacks draining protocols in seconds, human reaction times are increasingly insufficient. Agentic systems are now being deployed to monitor pending transactions, allowing for the activation of circuit breakers or the freezing of administrator keys within the same block window. Beyond security, these agents are streamlining compliance by automating real-time address screening and the generation of regulatory reports, tasks previously bogged down by manual data reconciliation.

However, this autonomy introduces significant vulnerabilities, including the potential for incorrect decisions or the manipulation of agents through prompt injection. CertiK stresses that because these systems lack legal personhood, accountability remains with the organizations and human supervisors. The firm warns that companies treating AI as standard software, rather than as a delegated member of the workforce, risk creating dangerous oversight gaps. To mitigate these risks, organizations must implement rigorous audit trails for every decision an agent makes, define strict authority limits, and maintain clear escalation protocols to ensure that human experts retain final control over high-stakes automated actions.