Agentic AI Risks Mount — Why Enterprises Must Secure Non-Human Identities Now

Okta announced its acquisition of AI security startup Permiso for approximately $200 million (TechCrunch). This deal marks a strategic pivot as enterprises scramble to secure the rising tide of autonomous agents operating across cloud environments.

Non-Human Identities Become the Primary Attack Surface

The shift toward agentic AI—AI systems capable of taking autonomous actions—has fundamentally altered the enterprise security landscape. Traditional security models focus on human users, but the emergence of AI agents introduces a massive influx of non-human identities (identities assigned to software processes or bots rather than people). These agents require specific permissions to access sensitive data, execute code, and interact with internal tools.

Security teams are now facing a mismatch where attackers move fluidly across cloud, identity, and device boundaries (SiliconAngle Tech). This fluidity makes traditional, isolated alert monitoring obsolete. Without a unified view of how an agent interacts with the broader infrastructure, a single compromised agent could serve as a permanent, undetected foothold for an attacker.

The risk profile changes the moment an agent moves beyond generating text to calling an API (The New Stack). When an agent has the authority to trigger a database query or initiate a fund transfer, it becomes a high-value target. Security professionals are now demanding strict permission boundaries to ensure these agents cannot exceed their intended functional scope.

Security Tooling Evolves to Combat Agentic Risks

The rise of autonomous software is forcing a rapid evolution in the cybersecurity stack. New specialized tools are entering the market to address the specific vulnerabilities inherent in agentic workflows. For example, Upwind Security Inc. recently unveiled its AI Agent Context Scanner (SiliconAngle Tech). This tool inspects the instructions, tools, and connections feeding an AI agent before those agents are permitted to act on them.

The complexity of these interactions requires a shift from reactive monitoring to proactive, real-time detection. Upwind Security Inc. has moved its AI Detection & Response (AI DR) service into general availability (SiliconAngle Tech). This service is designed to monitor agent behavior to identify deviations that suggest a hijacked or malfunctioning autonomous process.

For developers, this means that "linting"—the process of checking code for programmatic errors—is no longer sufficient for governance (The New Stack). Traditional code-checking methods cannot account for the unpredictable, non-deterministic nature of an AI agent making decisions in real-time. Instead, organizations must implement continuous monitoring of the agent's intent and its actual execution within the cloud environment.

Okta vs. The Emerging Agentic Threat Landscape

The Okta acquisition of Permiso highlights the massive scale of the identity threat detection market. By integrating Permiso's capabilities, Okta aims to provide visibility into the complex web of machine-to-machine interactions (TechCrunch). This is critical because, unlike humans, AI agents can operate at a scale and speed that makes manual oversight impossible.

The focus is shifting toward "attack path mapping," which uses graph databases to visualize how a threat could navigate through various cloud and identity layers to reach sensitive data (SiliconAngle Tech). This approach allows security teams to prioritize risks based on the actual potential for impact, rather than chasing isolated, low-priority alerts.

The Infrastructure Bottleneck: Compute and Data Control

As enterprises deploy more agents, the underlying infrastructure requirements are becoming increasingly complex. DataBahn recently raised $40 million to develop an "agentic data control plane" (SiliconAngle Tech). This platform sits between the systems generating enterprise telemetry (the data logs that record system activity) and the tools that consume it. It is designed to ingest, normalize, and route this data to ensure AI agents have the context they need without compromising security.

Simultaneously, the demand for high-performance compute is driving a consolidation of the AI stack. Nscale, a British AI neocloud, is acquiring Anyscale to gain greater control over the AI compute stack (TechCrunch). This move allows Nscale to better manage how workloads are scaled across diverse data centers and servers.

Infrastructure providers are also grappling with the tension between resource utilization and security. Protopia and Rafay are working to deliver multi-tenancy (the ability for a single software instance to serve multiple customers) for shared GPU AI factories (SiliconAngle Tech). This allows enterprises to convert idle GPU capacity into secure, token-metered services, ensuring that the high cost of compute does not lead to security compromises through resource isolation failures.

The Talent Gap in AI Implementation

Despite the influx of capital, the ability to deploy AI effectively remains a significant bottleneck for most organizations. A recent study estimates that only 2,000 U.S. engineers possess the specific expertise required to deliver meaningful AI ROI (TechCrunch). This scarcity is driving a desperate race among enterprises to hire forward-deployed engineers (engineers specialized in implementing complex software at a client site) to manage large-scale AI rollouts.

The complexity of the task is evident in the specialized nature of the tools being built. Developers are already creating local tools to manage multiple AI accounts and parallel agentic workflows (Hacker News). This grassroots development highlights a fundamental reality: the current enterprise tooling is not yet mature enough to handle the scale of agentic automation.

The goal for most organizations is to move from simple productivity gains to systemic, project-team outcomes (Certinia, via theCUBE Research). This requires more than just a few clever prompts; it requires a complete re-engineering of how software is developed, deployed, and secured.

Key Developments to Watch

• Okta (OKTA) (by end of 2025) — Integration of Permiso's non-human identity detection will be a key metric for their enterprise market share expansion.
• Nscale (Q3 2025) — The integration of Anyscale's scaling software will determine Nscale's ability to compete with major hyperscalers in the AI compute space.
• DataBahn (by November 2026) — The deployment of their agentic data control plane will signal whether enterprise telemetry can keep pace with agentic automation.