Codenotary flags 210,000 risky AI agent actions daily
Codenotary's AgentMon platform now monitors over 3 million AI-agent interactions daily across enterprise clients, flagging approximately 210,000—or 7%—as potentially unsafe or non-compliant, a signal that runtime security gaps in production AI systems are far more widespread than previously recognized.
According to the company, the vast majority of these anomalies stem not from external attackers exploiting vulnerabilities, but from the behavior of the AI agents themselves. The most common issues include the inadvertent exposure of passwords and API tokens, violations of pre-defined operational boundaries, and the onset of recursive workflows that can consume resources or lead to unpredictable outcomes. This finding challenges the traditional assumption that insider threats or malicious intrusions are the primary risk vectors for enterprise systems.
As AI systems transition from controlled pilot environments to live production, the report highlights that runtime risks are surging in parallel. Traditional security tools, which were designed to monitor endpoints and human user behavior, are poorly equipped to detect or govern the autonomous, often non-linear actions of AI agents. AgentMon addresses this blind spot by providing continuous observability across the entire stack, including agents, tools, APIs, and the data they interact with, offering enterprises a real-time view of agent behavior that conventional security frameworks miss.
The analysis points to a fundamental shift in enterprise infrastructure strategy: as AI moves from pilots to production, runtime governance is becoming foundational infrastructure, not an optional add-on. Enterprises that embed visibility and policy enforcement directly into their AI environments are succeeding in managing risk without slowing down the pace of deployment. The evidence suggests that proactive governance, rather than restrictive controls, enables organizations to maintain compliance and security while still capitalizing on the speed of AI innovation.
