Companies are finding agents where they’re not supposed to be at all. This was to be expected, as agentic AI has been adopted across the entire development pipeline. Palo Alto Networks’ 2026 State of Cloud Security Report reveals that 73 percent of respondents feel comfortable letting agents rewrite code, adjust configurations in production environments, and run security tests.
This is Palo Alto Networks’ sixth State of Cloud Security Report, which now includes 2,800 surveyed security decision-makers and staff across 10 countries. The message from this group of respondents is clear: agents are everywhere. A staggering 99 percent deploy AI agents for (semi-)autonomous actions in the development pipeline.
Alarming pace
The pace may sound alarming to a CISO who hasn’t approved such a high adoption rate at all, but it shows just how inseparable AI has become from developers in 2026. There are results to show for it, though it’s difficult to determine exactly how much more productive staff actually are thanks to the technology. According to 82 percent, AI is at least resulting in more infrastructure updates being deployed to the cloud than before.
At the same time, organizations enforce, on average, fewer than three out of eight measured pre-deployment checks to roll out those updates, a sign that agents aren’t entirely compatible with legacy security techniques. In 42 percent of cases, AI generates infrastructure-as-code templates, but only 34 percent scan that code before production.
On average, 19 percent of severe and critical vulnerabilities made it into production within 90 days. Implementing a fix takes an average of 18.4 days, which gives attackers ample time to exploit a vulnerability if it is also detectable via the internet. Business pressure to release updates quickly is the most commonly cited reason for this high figure (39 percent).
Human or agent?
Accountability is also lacking, according to the report. Two in five organizations cannot reconstruct what an agent has done. In the financial sector, that figure rises to 46 percent. In 53 percent of cases, agent activity is logged under a developer’s identity or personal access token. In the event of an incident, investigators must therefore first determine what a human did and what an agent carried out. Ultimately, the responsibility lies with the human, but for analysts and for improving AI tools, this distinction is necessary. In addition, revoking access for non-human identities still requires manual intervention in 60 percent of cases. We previously wrote about why AI agents make the perfect insider.
Remarkably, confidence remains high: 97 percent believe they can secure the AI supply chain. Only 30 percent can detect prompt injection.
Mythos-grade threats
Virtually all respondents (98 percent) are concerned about frontier models that independently discover and exploit vulnerabilities. The survey refers to “Mythos-grade” threats, a reference to Anthropic’s Claude Mythos. Palo Alto was a launch partner for Project Glasswing, which gave security researchers early access to the Claude Mythos Preview in April. According to Anthropic, that model discovered thousands of unknown vulnerabilities in operating systems and browsers.
The report recommends giving agents their own identity, with a clear owner and minimal privileges. Palo Alto itself sells tools for this purpose. Prisma AIRS secures AI agents, while Cortex AgentiX deploys agents for security operations.