Incident sheet · /security

Filed
Read
4 min · 873 words
Severity
Routine

MCP Security Tools Lead an Uneven Agentic AI Market

Nemertes Research says MCP gateways and registries are likelier to work reliably, while semantic firewalls remain early-stage and require testing in defined enterprise use cases.

CISO’s guide to agentic AI runtime security and governance
FeedCISO’s guide to agentic AI runtime security and governance — Tho-Ge / pixabay

Event log

  1. MCP gateways and registries are identified as the new agentic AI security tools most likely to work reliably.
  2. The guide identifies three identity layers: agent-specific, orchestrator and end-user IDs.
  3. Semantic firewalls can block all prompt-injection attacks and certain data-poisoning attacks, but remain early-stage.
  4. SIEM and user entity behavioral analytics systems may lack capacity or speed for monitoring swarms of ephemeral agents.
  5. John Burke joined Nemertes Research in 2005 with nearly two decades of technology experience.

By Daniel Okafor

Model Context Protocol (MCP) gateways and registries are the new agentic AI security tools most likely to work reliably, according to John Burke, CTO and research analyst at Nemertes Research. The guide calls for controlled testing before CISOs accept vendor claims. Burke's assessment was blunt: “Your mileage may vary.”

What must CISOs secure first?

Enterprises are deploying agents to pursue AI returns, often without securing the agents or wider business against them. The Nemertes guidance says protection and control can be one and the same capability. CISOs should extend zero-trust architecture rather than treat agent security as a separate product category. It assumes organizations should already be building a comprehensive zero-trust environment.

The hard part is identity. The guide says multi-tiered identities may ultimately be required, combining agent-specific, orchestrator and end-user IDs so policy can enforce least-privilege access. Yet it does not present these as a finished standard: questions about trust and identity remain difficult for systems whose actions may span agents, orchestrators and end users.

Which existing controls still work?

The Nemertes guide assigns established controls specific roles:

  • Identity and access. The guide says digital identities let identity- and role-based controls limit agent access. IAM may supervise persistent personal or departmental assistants that orchestrate shorter-lived agents. Systems built for humans and long-term service accounts fit ephemeral agents poorly, though container or microservice identity systems may adapt in some environments.
  • Data loss prevention. The guide calls DLP a necessary backstop because in-agent controls are prone to failure.
  • API and browser enforcement. The guide identifies API gateways as policy controls for resource access. Secure enterprise browser management restricts agents from unsanctioned websites.
  • Monitoring. The guide says some SIEM and user entity and behavioral analytics systems can track agent activity and trigger alerts. Yet many lack the capacity or speed required for monitoring floods from swarms of ephemeral agents, while others can absorb the flood at too high a cost.

What do the new tools add?

The guide positions these controls as a combined layer for enterprise agents: semantic intent, path history and explicit governance of MCP servers and their sanctioned functions.

  • Semantic firewalls. The guide says these tools monitor prompts and data entering an agent, plus the responses it generates. They can block all prompt-injection attacks and certain data-poisoning attacks. On output, they enforce DLP rules and detect model drift as responses become less reliable. They consider meaning and intent alongside data content and network address.
  • Path-dependent access management. The guide describes this established application method as tracking whether an agent passed required states before requesting access. Its example is straightforward: no password-change page before identity-verification page. New monitoring and analytics tools follow agents across their lifecycle and use history in policy decisions.
  • Path caveat. The guide says agent paths can be nondeterministic, include loops or recursion, and the monitoring tools are still evolving.
  • MCP registries. The guide says they maintain an inventory of intended servers and sanctioned functions.
  • MCP gateways. Adapted from API gateways, they monitor and enforce policy for MCP calls and responses. They can redact protected information, block responses containing it, and filter prompts directly or through semantic firewalls.

The guide calls registries and gateways among the most important tools because they focus on MCP, a major AI integration standard. MCP remains relatively young, and implementations are maturing.

Companies may face a steady stream of server vulnerabilities requiring patches or workarounds. They may also run rogue MCP servers or leave obsolete ones active. Registries can address inventory, but neither tool removes that operational burden.

Are these tools worth the cost?

The guide offers no universal winner. CISOs and security teams should test each tool in well-defined use cases, according to Nemertes. Each test should establish:

  • Whether promised controls work.
  • Whether staff effort remains reasonable.
  • Whether results justify the tool's cost.

The strongest recommendation is procedural rather than categorical: run defined use-case tests, measure staff effort and confirm that each promised control delivers its claimed value.

The guide says MCP gateways and registries resemble existing gateway and registry methods, making them the most likely new tools to work reliably. Emerging approaches are less likely to deliver value to everyone who tries them, according to Nemertes.

The guide says semantic firewalls resemble web application firewalls. To be more than an expensive WAF, they must apply natural language processing at speed. They could deliver significant value in some use cases. For now, they remain in their early days. Their business case therefore depends on tests showing specific, reliable gains.

What should CISOs test first?

The guide's sequence starts with established controls:

  • Digital identity and least-privilege policy.
  • DLP backstops.
  • API enforcement.
  • Runtime monitoring.

AI-specific controls should earn their place through defined tests, according to Nemertes. No vendor category receives a blanket recommendation.

Burke joined Nemertes in 2005, bringing nearly two decades of technology experience. His background spans end-user support, programming, systems and database administration, network administration and architecture, and systems architecture. That breadth does not validate any vendor. It does frame the central test: an agentic control must work inside the enterprise, under real policy, with reasonable staff effort.

via techtarget.com (Source)

Filed under

  • model-context-protocol-mcp
  • agentic-ai-security
  • zero-trust-architecture
  • identity-and-access-management
  • mcp-gateways
Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Staff writer covering technology and product news in enterprise information technology and the leaders who run at CIO Monitor.

203 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›