Security leadership is being pulled in three directions at once as enterprises move deeper into production AI and expand application footprints across hybrid cloud environments. A new application strategy survey from F5 highlights how distributed AI inference is becoming an operational priority, how attacker tactics are evolving toward faster and more automated exploitation, and why hybrid multicloud is increasingly a permanent operating model rather than a temporary transition.
For CISOs and security leaders, the implication is straightforward but complex to execute: protecting AI applications now requires governance and controls at the inference layer, not only around traditional perimeter defenses or model development stages. At the same time, organizations need consistent visibility and policy enforcement across environments that are inherently distributed, heterogeneous, and increasingly tied together by APIs, identity flows, and runtime data movement.
Distributed AI inference is moving to the center of enterprise operations
Historically, enterprise AI programs often emphasized model training, tuning, and experimentation. The latest survey findings suggest that the center of gravity has shifted. Respondents reported that inference, rather than training or tuning, is the dominant AI activity. Inference workloads are also described as distributed, running across data centers, cloud environments, and edge locations.
In practice, distributed inference changes what security teams need to monitor and control. Inference is frequently delivered through endpoints and APIs, with prompts and context passed between services and model components. When multiple models are involved, organizations may route requests through chained flows or dynamically select models, creating additional runtime complexity and new potential failure or abuse points.
Why inference-layer protection is different
Security challenges cited in the survey include the sprawl and management overhead that comes with operating multiple models. Chained routing can introduce risks at runtime because prompts, responses, and contextual data move between models. Additionally, organizations are expected to maintain performance, security, and reliability simultaneously, which requires end-to-end observability. The goal is not just to detect issues, but to ensure that security and compliance policies apply consistently regardless of where a model runs or how a request is routed.
There is also a cost and operations angle. High-performing inference infrastructure can be expensive to operate, and security tooling must be integrated in a way that does not add uncontrolled latency or operational burden. The survey frames distributed inference as an application tier in its own right, suggesting security leaders should treat it with the same operational discipline applied to other critical application components.
Threats are accelerating as AI becomes part of both offense and defense
Alongside the rise of AI inference, the survey points to a rapidly changing threat landscape. Attackers are described as using AI to accelerate discovery, automate execution, and extend their reach. In parallel, enterprises are adopting agentic AI, which increases the number of actions an AI system can take and expands the set of permissions and access pathways it may require.
The report indicates that most organizations are modifying applications to enable autonomous AI agent access. It also reports an expected gap around identity and access control for these agents. If AI agents require permissions to act across systems, then traditional identity governance becomes a more central security problem, especially when access tokens, session context, and authorization decisions are generated or used dynamically as part of AI workflows.
New attack patterns and runtime risks
Security concerns associated with AI systems frequently include prompt injection, data poisoning, and model inversion. While these terms have been discussed broadly across the industry, the core issue for security leadership is how they map to practical controls in live systems. AI does not only add new data types and endpoints, it also changes how interactions occur at runtime, turning prompts and model outputs into elements that can be manipulated or used as part of an attack chain.
As a result, the survey argues that tools need to focus on the inference layer where APIs, prompts, and access tokens intersect in real time. It also emphasizes the importance of monitoring and inspection of prompts and responses, with visibility integrated into delivery and security workflows to reduce gaps that can be exploited.
Hybrid multicloud is now the baseline, not a temporary phase
The third shift in the survey is operational: hybrid multicloud is increasingly described as the default environment for most enterprises. The survey data indicates that the majority of respondents manage multicloud environments and that application deployments commonly span on premises, public cloud, and co-location facilities. It also suggests that enterprises may manage multiple data centers, multiple co-location sites, and several public cloud providers.
For security teams, distributed infrastructure amplifies the complexity of enforcing consistent controls. Security policies that work in one environment may not map cleanly to another due to differences in network topology, orchestration patterns, identity integrations, and runtime behavior. With distributed AI inference relying on APIs and application workflows, the security challenge becomes less about protecting isolated segments and more about ensuring consistent enforcement across the end-to-end path of an inference request.
Consistency across teams and control planes
The survey frames this as an argument for a more unified approach to delivering and securing applications. It suggests collapsing fragmented control points into a single enforcement layer to address gaps created by dynamic AI systems and hybrid multicloud architectures. From a leadership perspective, the operational goal is to align NetOps, SecOps, and application teams around a shared operational model, rather than managing security rules and observability across separate tooling and control planes.
While the survey is vendor-led and includes references to specific product positioning, the underlying operational themes align with broader industry concerns: inference and agentic workflows require runtime controls, hybrid deployments increase the difficulty of consistent policy enforcement, and security monitoring must span the full request and response lifecycle.
What security leaders can take away
Even without adopting a specific platform strategy, the survey points to concrete planning priorities for enterprise security leaders:
- Operationalize inference security by treating inference as a first-class application tier with clear monitoring and policy enforcement requirements.
- Reassess identity and access for AI agents to ensure authorization controls remain effective when AI systems make requests and act across services.
- Extend observability to prompts and outputs where appropriate, so security teams can evaluate real-time behavior and detect abnormal patterns in live workflows.
- Design for hybrid consistency so that security controls apply reliably across on premises, public cloud, and co-location environments.
- Reduce fragmented enforcement to minimize policy gaps that can emerge across complex routing and multi-model interactions.
The shift to distributed AI inference is already reshaping application architectures. Meanwhile, a threat environment that is faster and more automated, combined with hybrid multicloud reality, is raising the bar for how security leaders plan governance, instrumentation, and enforcement. The organizations that handle these changes effectively are likely to be those that treat AI inference and agent workflows as operational workloads, with security built into the same discipline as performance and reliability.
