40% of organizations struggle to retrace AI agents’ actions
A significant portion of organizations struggle to trace the actions of AI agents in cloud environments, impacting security oversight.

Forty percent of organizations report challenges in reconstructing the actions of AI agents, according to a report from Palo Alto Networks. This issue is particularly pronounced in the financial sector, where the percentage rises to 46%. The findings highlight the growing complexity of managing AI-driven development in cloud infrastructure.
Palo Alto Networks' State of Cloud Security Report 2026 reveals that 99% of organizations utilize AI agents for autonomous or semi-autonomous actions within their development pipelines. While AI-driven development has reportedly increased the frequency of updates to cloud infrastructure, it has also introduced significant security challenges. Nearly three-quarters of organizations allow these agents to modify code and configuration settings directly in production environments, which raises concerns about the oversight of these changes.
The report indicates that the actions of AI agents can often be approved before understanding their potential security risks. In many cases, reviewers may authorize changes without fully inspecting the dependencies, secrets, images, or infrastructure involved. Only about one-third of organizations have the capability to trace the journey of code from its initial commit to its execution in the cloud, and many do not enforce manual security reviews consistently.
Infrastructure as code is a practice that allows developers to manage and provision cloud resources through code, which can speed up deployment processes. However, this approach also increases the risk of misconfigurations reaching production environments before security teams can identify and rectify them. The report highlights that organizations typically enforce fewer than three of the eight essential predeployment security checks, which can leave significant gaps in their security posture.
The report shows that, on average, 19% of high- and critical-severity application security issues made it into production within the last 90 days. Organizations also take an average of 18.4 days to deploy fixes after discovering these security issues. The pressure to release software quickly is cited as a barrier to effective predeployment remediation by many organizations, with business targets and tight deadlines often overshadowing security considerations.
Concerns about undefined risk acceptance also pose challenges for organizations, as 30% report that the lack of explicit criteria allows known vulnerabilities to reach production without clear ownership or timelines for resolution. This issue is particularly prevalent in Germany and India, where a higher percentage of organizations cite it as a significant barrier.
The report reveals that 40% of organizations have been unable to trace the actions of AI agents, with the financial sector facing the highest difficulty. Investigators often struggle to differentiate between actions taken by human developers and those executed by AI agents, as many engineers log agent activities under their own identities. This practice complicates incident response efforts, particularly when credentials need to be revoked or replaced.
Despite the challenges, 97% of organizations express some level of confidence in their ability to implement a security strategy for their AI supply chain and applications. However, only 36% can effectively monitor AI behavior in production, and a mere 30% can detect prompt injection attacks. As AI agents increasingly manage cloud resources, security teams must prioritize the detection of unauthorized actions and ensure that agents operate within defined access parameters.
What it meansOrganizations face significant difficulties in tracking AI agent activities, impacting their ability to manage security risks effectively.
Ask your IT team this weekHow do we currently log and monitor actions taken by AI agents in our environment?What measures are in place to ensure that AI agents operate within defined security parameters?
DecisionThe IT team reviews logging practices for AI agent activities, and security checks that improvements are made.