Agentic AI is pouring into enterprise IT environments like floodwaters into a basement.
Enterprises, racing to realize ROI from AI, often deploy these agents without considering how to secure them or how to secure the business against them. CISOs need to step up and revise their cybersecurity architectures with capabilities that both protect and control agents — which can be one and the same.
Effective agentic AI runtime security requires CISOs to fold them into a broader zero-trust architecture. Yes, the organization should already be building a comprehensive zero-trust environment. And yes, there will be some difficult questions about what trust means in the context of agentic systems that could ultimately require the development of multi-tiered identities for agents — identities that combine agent-specific IDs with orchestrator IDs with end-user IDs to allow enforcement of least-privilege access.
In short, CISOs must blend all the familiar components of a modern cybersecurity architecture with tools specifically designed to manage the problems AI generates.
Focus on cybersecurity fundamentals
A strong enterprise cybersecurity environment should already have a broad array of tools in place to address the challenges posed by agentic AI.
For example, most environments already use identity- and role-based access controls throughout their infrastructure. They also try to enforce the principle of least access to and among enterprise resources. To that end, a key requirement for securing agents and the environments they work in is to give those agents digital identities. Once that happens, existing identity-based controls can limit agent access in accordance with defined policy.
Identity and access management (IAM) tools can also help oversee some agents, such as persistent personal or departmental assistants that act in part as orchestrators of other, more ephemeral agents. The shorter-lived the agent, however, the less likely existing IAM platforms — built for humans and long-term service accounts — will work. Systems built to handle identity for containers or microservices, however, could be adapted to meet these needs in some environments.
Likewise, to ensure that protected information is managed correctly, existing tools that detect and prevent data loss provide useful backstops to in-agent controls. These controls are prone to failure, so data loss and prevention (DLP) backstops are necessary.
API gateways offer another policy-enforcement method for controlling agent access to other resources, while some secure enterprise browser management tools restrict the use of AI agents within browsers, making it more difficult for agents to go to unsanctioned websites.
Lastly, some existing SIEM and user entity and behavioral analytics systems can be extended to serve as a focal point to track agent activity and trigger alerts. Yet, many do not have the capacity nor speed required to handle a flood of monitoring data from swarms of ephemeral task-oriented agents. Others might be able to handle the flood, but they will cost too much to use in that capacity.
AI-specific tools for agentic AI runtime security
Vendors have responded with a new generation of AI-focused tools. The segment is growing fast and already includes such novel approaches as semantic firewalls and path-dependent access management. These make great additions to a zero-trust architecture because, in answering the question of whether an identity is allowed to do what it is trying to do, they consider meaning and intent as well as nuts and bolts questions of data contents and network addresses. Additional approaches include Model Content Protocol registries and MCP gateways.
Semantic firewalls are evolving rapidly but can already be useful. They monitor what prompts and data are fed into an agent at runtime, as well as the responses it generates. On the input side, they spot and block all kinds of prompt injection attacks and certain data poisoning attacks. On the output side, they are part DLP enforcer by preventing agents from sending data they are not supposed to send, and part model drift detector, watching for models that gradually becoming less reliable in their responses.
Path-dependent security is well understood in application development: It prevents a process from reaching a particular state if it has not passed through other required predecessor states — for example, no access to the password change page without first visiting the identity verification page. For agentic AI runtime security, tracking the path an agent has taken to get to the point where it's requesting access is a crucial determinant of whether to grant that access. New monitoring and analytics systems track these paths. They track agents across their lifecycle in an environment, allowing policy enforcement to take agents' histories into consideration. This is a complex challenge at this level. Agent action paths can be non-deterministic and include looping and even recursion; the tools to monitor these behaviors are still evolving.
MCP registries and MCP gateways are among the most important, as each focuses on AI's major integration standards. MCP is relatively young, and implementations are maturing. As a result, companies could face a steady stream of server vulnerabilities that have to be patched or worked around. At the same time, people might spin up rogue MCP servers or leave obsolete ones running instead of properly decommissioning them. MCP registries help ensure an environment has only the MCP servers it means to have and that the sanctioned functions of each are understood. MCP gateways, adapted from API gateways, serve as both monitoring and policy enforcement points for MCP calls and responses. They also perform DLP-style functions, redacting protected information from responses or blocking responses that contain it. They can also perform prompt detection and filtering, either independently or via semantic firewalls.
These new tools. Are they worth it?
"Your mileage may vary" is the only possible answer. CISOs and their teams need to test specific tools in well-defined use cases to ensure they deliver on their promises with a reasonable amount of staff effort.
Tools similar to existing methods, such as MCP gateways and registries, are most likely to be effective and reliable. Tools that are just emerging are less likely to deliver value to everyone who tries them. Semantic firewalls, for example, look like a twist on web application firewalls, but to be anything more than an expensive WAF, they have to break new ground in applying natural language processing at speed. They could provide significant value in some use cases, but for now, they remain in their early days.
John Burke is CTO and a research analyst at Nemertes Research. Burke joined Nemertes in 2005 with nearly two decades of technology experience. He has worked at all levels of IT, including as an end-user support specialist, programmer, system administrator, database specialist, network administrator, network architect and systems architect.