Security Teams Cannot Track What AI Coding Agents Do, Palo Alto Networks Finds
Sixth annual survey of 2,800 security leaders finds governance gaps widen as agent autonomy grows

Palo Alto Networks published its sixth annual State of Cloud Security Report on November 16, 2026, finding that organizations have deployed AI coding agents at near-universal scale while losing the ability to audit what those agents actually do. Among 2,800 security decision-makers surveyed across ten countries, 99 percent now run AI agents in their development pipelines. Forty percent cannot reconstruct the sequence of actions an agent took after the fact — rising to 46 percent among financial services firms. Those same organizations report that 73 percent are comfortable allowing agents to rewrite code and modify production configurations. The combination describes a governance posture that practitioners have begun calling accountability-free autonomy.
The report arrives as organizations face a structural identity design failure that monitoring tools alone cannot address. When engineers deploy AI coding assistants — through IDE extensions, locally hosted MCP servers, or API integrations — those tools typically authenticate using the developer's existing personal access token, or PAT. PATs were designed for short-duration automation by a single person. They are now serving as the primary authentication credential for multi-step autonomous agents that run continuously, execute hundreds of API calls, write directly to repositories, and modify infrastructure configuration. The result is that 53 percent of agent activity in surveyed organizations is logged under a human developer's identity. When an audit trail is needed, it cannot distinguish between what a developer did and what an autonomous system did on their behalf.
Agents Have Inherited Enterprise Credentials They Were Never Designed to Hold
The scale of the credential mismatch is more severe than most organizations recognize. A companion Palo Alto Networks 2026 Identity Security Landscape Report found that machines now outnumber human identities within surveyed enterprises by a ratio of 109 to 1. That disparity did not exist three years ago, before widespread AI coding tool adoption created a new category of persistent machine actor operating inside the enterprise perimeter under borrowed human credentials.
The problem compounds at the lifecycle management layer. Sixty percent of surveyed organizations require manual intervention to revoke non-human identity access when an agent is decommissioned, a developer leaves, or a project ends. In a world where machine identities are provisioned faster than human identities by an order of magnitude, manual revocation creates a standing inventory of orphaned credentials — access tokens that remain valid long after the agent or workflow that used them has stopped. Any attacker who obtains one of those tokens inherits whatever permissions the developer held when they originally set it up.
The credential problem is worsened by the emergence of Model Context Protocol servers as an ungoverned attack surface. MCP servers are locally or remotely hosted services that expose tools and data sources to AI agents over a standardized protocol. A developer can install an MCP server that connects their coding agent to a production database, an internal API, or a file system directory — typically without any IT governance review. The Cloud Security Alliance documented this dynamic in its AI Velocity Gap research, finding that 68 percent of organizations claim high confidence in their visibility into developer AI tools, yet 82 percent have already discovered AI agents operating in their environment that they did not know existed.
Deployment Moves Faster Than Pre-Production Security Checks
The velocity gap extends into the production pipeline. Palo Alto Networks found that organizations enforce fewer than three of eight measured pre-deployment security checks on AI-generated code and configuration templates. Infrastructure-as-Code templates — the files that provision cloud resources, configure network rules, and assign identity access policies — are a particular weakness. Only 34 percent of organizations scan AI-generated IaC templates before they reach production. Forty-two percent never scan them at all.
The consequence is measurable. Nineteen percent of severe and critical vulnerabilities introduced through AI-generated code reach production systems within 90 days of introduction. The mean time to remediate once discovered is 18.4 days. For organizations where AI agents are generating the infrastructure configuration that governs their security posture, that gap between introduction and remediation is the window in which a misconfigured storage bucket, an overprivileged IAM role, or an exposed network rule exists in production.
The report finds that 100 percent of surveyed organizations reported a major cloud security incident in the past 12 months — a figure that reflects both the operational scale of cloud environments and the increasing role of AI-generated code in those environments.
The Confidence Gap Separates Belief From Capability
The most structurally significant finding in the report is not any individual failure metric but the gap between confidence and capability. Ninety-seven percent of surveyed organizations say they are confident in their ability to secure the AI supply chain. Only 30 percent of those organizations can actually detect prompt injection attacks — the technique by which an attacker embeds malicious instructions in content that an AI agent processes as ordinary input, causing the agent to execute those instructions as if they originated from its principal. A prompt injection in a code comment, a document an agent reads, or an API response an agent queries is effectively an unauthorized lateral entry point that bypasses every traditional access control.
The confidence gap is not unique to Palo Alto Networks' data. The Cloud Security Alliance found independently that 57 percent of employees use AI coding tools without formal IT approval, and 68 percent of organizations lack visibility into the full scope of developer AI tool deployment. Arkose Labs, in its March 2026 Agentic AI Security Report covering 300 enterprise security leaders, found that 97 percent expect a material AI-agent-driven security or fraud incident within the next twelve months — a figure that sits in direct tension with the 97 percent confidence figure Palo Alto Networks documented. Roey Eliyahu, chief executive of Salt Security, described the situation facing organizations that have discovered more agents than they authorized: "That is not a visibility problem. That is a governance crisis."
What Frontier-Model Access Changes for Enterprise Threat Modeling
The report introduces a threat category that did not feature prominently in previous annual editions: the concern that frontier AI models will independently discover and exploit enterprise vulnerabilities. Ninety-eight percent of surveyed organizations express concern about this scenario. The concern is no longer theoretical. As of July 2026, Anthropic restored access to Claude Fable 5.2 and Claude Mythos 5.2 following a brief export control suspension by the U.S. Department of Commerce, and both models are available to enterprise customers with established security programs.
Palo Alto Networks was identified as a launch partner for Anthropic's Project Glasswing, the program through which a small number of organizations received early access to Claude Mythos capabilities for security evaluation. The company's access to frontier models for defensive security research provides context for the report's framing of "Mythos-grade threats" — autonomous model capabilities sophisticated enough to chain vulnerability discovery with exploitation without requiring human orchestration. Organizations that cannot reconstruct what an AI coding agent did in their environment are not positioned to detect or respond to that threat class.
How Organizations Close the Accountability Gap — and What Palo Alto Networks Sells
A set of technical principles has emerged for constraining agent autonomy without eliminating its usefulness. Meta formalized the Rule of Two in an October 2025 blog post that has since been adopted as a baseline design principle by several cloud providers and enterprise security frameworks. The rule holds that agents should combine no more than two of the following three properties: processing untrusted input, having access to sensitive data, and having the ability to change state or communicate externally. Any agent that holds all three has no natural containment perimeter. An agent that checks a public repository and writes a summary document is bounded. An agent that processes pull-request comments, reads production database credentials, and can push commits is not.
The governance category that addresses these problems is called non-human identity management. Where human identity governance has decades of tooling, standards, and regulatory precedent behind it, NHI governance is an emerging discipline. Its core requirements are machine identity inventory, access lifecycle management, behavioral monitoring, and attribution — the ability to associate an action with the specific agent that took it rather than the human whose token it used.
Palo Alto Networks sells products that directly address every gap described in its own report. Prisma AIRS 3.0, relaunched at RSA Conference in March 2026, is the company's AI security posture management platform; its AI Agent Gateway provides a central control plane for governing which agents can connect to which systems and monitoring agent behavior. Cortex Agentic Endpoint Security, which emerged from the company's February 2026 announced intent to acquire Koi Security, closes the endpoint blind spot for agents running outside cloud-managed environments. Project Perception, expanded in October 2026, governs what the company calls "agentic insiders" — agents operating with internal access that behaves like a privileged insider account without the accountability structures applied to human employees.
The vendor conflict of interest here is real and worth naming explicitly. Palo Alto Networks commissioned and published a survey documenting governance failures that its own commercial products are designed to remediate. The finding should be read as a data point within that context. That said, the directional findings are corroborated by independent research from the Cloud Security Alliance, Arkose Labs, AvePoint, Gravitee, and NeuralTrust, all of which documented overlapping governance failures in 2026 without a commercial interest in any particular remediation product. The underlying problem the report describes — organizations deploying autonomous systems faster than the governance frameworks required to secure them are emerging — is consistent across the independent data.
The Accountability Gap Will Widen Before It Narrows
The deployment trajectory suggests the gap will grow before it closes. Gartner estimates that by the end of 2026, 40 percent of enterprise applications will incorporate task-specific AI agents, up from fewer than five percent in 2025. That eightfold increase in agent density across enterprise environments, combined with the credential inheritance and visibility failures the Palo Alto Networks report documents, means the pool of unaudited agent actions will expand substantially before the governance infrastructure to constrain it matures.
The practical stakes are illustrated by a 2026 incident at Hugging Face, in which a compromised AI agent executed 17,600 discrete actions over 4.5 days before it was detected. That incident did not require a novel attack technique. It required only that an agent with broad credentials and no behavioral monitoring be given a task it could accomplish autonomously.
For the 40 percent of organizations that cannot reconstruct what their AI coding agents have done, the accountability gap is not a future risk. It already exists. Regulated industries face an additional structural exposure: an organization that cannot produce an audit trail of AI agent actions in a system covered by SOX, DORA, or comparable financial services regulation cannot demonstrate that its access controls functioned as intended — regardless of what those controls say on paper. The question the next twelve months will answer is whether that exposure generates meaningful regulatory consequence before it generates a material incident, or the other way around.



