Palo Alto Networks finds 40% of surveyed organizations cannot retrace AI agents’ actions
The survey of 2,800 security professionals finds agents working under developers’ identities, limited monitoring and slow responses when credentials are compromised.
Palo Alto Networks’ State of Cloud Security Report 2026, published November 16, finds that confidence in securing AI systems is far ahead of operational visibility: 97% of respondents expressed some confidence, but only 36% monitor AI behavior in production. The survey also found that 53% record agent activity under a developer’s identity, complicating incident investigations, and 60% still need manual intervention to revoke nonhuman access. The findings point to identity, monitoring and recovery controls as practical gaps for teams deploying agents.
01
The survey covered 2,800 security executives and practitioners in 10 countries; its results are respondents’ accounts, not an independent systems audit.
02
Forty-three percent of organizations found agents or automated workflows operating without security approval or oversight in the prior year.
03
Organizations enforce fewer than three of eight measured predeployment checks on average, while 42% use AI to generate infrastructure-as-code templates and 34% scan them before production.
AI agents are getting permission to change live software systems without leaving a trail that organizations can reliably reconstruct. Palo Alto Networks’ newly published State of Cloud Security Report 2026 finds that 73% of surveyed organizations allow agents to modify code, deployment settings or production environments, while 40% have been unable to retrace an agent’s actions.
Published November 16, the report draws on a survey of 2,800 security executives and practitioners in 10 countries, according to SiliconANGLE. Its findings describe respondents’ accounts of their organizations, rather than an independent audit of their systems. Within that sample, 99% said they use agents for autonomous or partly autonomous work in software development pipelines—the processes that move software from code changes to deployment.
When the agent wears a developer’s identity
One source of confusion is how agents identify themselves. In 53% of surveyed organizations, agent activity is recorded under a developer’s identity or personal access token—a credential used to access systems. Those records blur the distinction between a person’s work and an agent’s work. Investigators must separate the two before they can determine what an incident affected.
The tracing problem was most common in financial services, where 46% said they could not reconstruct an agent’s actions. Oversight gaps also extend beyond recordkeeping: 43% of organizations found AI agents or automated workflows operating without security approval or oversight during the preceding year, Help Net Security reported.
Cutting off access is not necessarily straightforward. Sixty percent still require manual intervention to revoke access for nonhuman identities. Organizations take an average of 2.5 days to revoke or replace compromised developer credentials, API keys or pipeline secrets. Revoking a shared credential can also disrupt services that depend on it, requiring security and engineering teams to coordinate replacement access.
Confidence outpaces monitoring
97%Confidence in AI security
Respondents at least somewhat confident in their ability to secure their AI supply chain and applications.
36%Monitoring in production
Organizations able to monitor AI behavior in production, where deployed systems do their work.
Editorial illustration for Palo Alto Networks finds 40% of surveyed organizations cannot retrace AI agents’ actions.Source: helpnetsecurity.com.
More changes, limited checks
Agents are also increasing the amount of work security teams must inspect. Eighty-two percent of respondents said AI-driven development increased the volume of cloud infrastructure updates. Yet organizations enforce fewer than three of the eight predeployment security checks measured in the survey, on average. Only 33% can trace code from its recorded change through to its operation in the cloud across most production workloads.
A specific gap concerns infrastructure as code: templates that define cloud resources and settings. Forty-two percent use AI to generate those templates, but only 34% scan that code before production. Faster provisioning can therefore carry configuration mistakes into cloud environments before security teams detect and address them, particularly when checks before deployment cannot keep pace.
Respondents said an average of 19% of high- and critical-severity application security issues reached production over the preceding 90 days. Getting a fix into production after identifying a code security issue takes an average of 18.4 days. These figures cover application security generally; they are not a count of vulnerabilities caused by agents.
The obstacles are partly organizational. Business pressure to release quickly hinders fixes before deployment at 39% of organizations, while 35% cite concerns about slowing development. Undefined risk acceptance is a barrier for 30%: known weaknesses can ship without clear agreement about who owns the exposure or when a fix is due.
Finding the activity, then assigning ownership
Data oversight has similar blind spots. Only 9% can automatically discover and classify ungoverned AI activity across cloud environments, while 70% cannot map sensitive-data flows across AI systems and workflows. Security teams also lack a joined-up view: just 19% report unified monitoring data and workflows across code, cloud, identity and running applications. Forty-seven percent manually connect application vulnerabilities with threats to running systems.
The report’s recommendations put ownership and recovery alongside automation:
Give agents their own identities, a clear owner and only the permissions needed for their jobs.
Automate fixes for recurring issues only after establishing procedures to roll changes back.
Sources
siliconangle.comPalo Alto Networks report finds organizations struggle to track and secure AI agents in the cloud - SiliconANGLE
helpnetsecurity.com40% of organizations struggle to retrace AI agents’ actions - Help Net Security
Reader comments
Newest comments first. Replies stay oldest first.