cd /news/ai-agents/building-a-security-framework-for-th… · home › topics › ai-agents › article
[ARTICLE · art-171278] src=blogs.cisco.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

Building a Security Framework for the Agentic Enterprise

Cisco researchers Flemming Andreasen and a co-author published a vendor-neutral whitepaper, "Identity-Centric Security for the Agentic Enterprise," proposing a reference architecture that carries agent identity, delegation and policy across runtimes, identity services, networks, APIs and Model Context Protocol (MCP) servers. A Cisco survey of 1,380 IT and cybersecurity leaders found 24% reported an incident in the prior 12 months involving an AI agent or other non-human identity with compromised, excessive or unauthorized access, while only 9% had fully integrated AI-agent and non-human-identity activity into security detection and response, 42% partially and 39% not at all. The framework follows the MCP project's July 2026 specification update, which names authorization as a substantial source of implementation effort and hardens the flow with short-lived, narrowly scoped access tokens tied to the delegating human or organization.

by read3 min views1 publishedNov 17, 2026
Building a Security Framework for the Agentic Enterprise
Image: Blogs (auto-discovered)

When an AI agent takes an action, a familiar identity question becomes more complicated: who is acting, on whose authority, against which resource, and how can we reconstruct what happened?

A person may delegate a task; the agent may then select tools, call services or other agents, and move data through several systems – even through the internet. If identity and policy context stop at a gateway or disappear between calls, the enterprise loses the chain of accountability it needs to manage risk.

That makes agent security an ecosystem problem. Existing controls remain necessary, but they must work together across runtimes, identity services, networks, APIs, Model Context Protocol (MCP) servers, other agents and security operations. Security teams need a consistent way to identify managed agents, discover activity and resources they do not yet manage, scope authority to each task, enforce policy where actions occur, and connect activity to detection and response. Individual controls can address parts of the problem; an end-to-end architecture shows how context moves between the ecosystem and how decisions are enforced.

The gap is visible in the data. In a Cisco survey of 1,380 IT and cybersecurity leaders, 24% reported an incident in the prior 12 months involving an AI agent or other non-human identity with compromised, excessive, or unauthorized access.

The survey also found that 9% had fully integrated AI-agent and non-human-identity activity into security detection and response; 42% said integration was partial and 39% said there was none.

Protocol maintainers are seeing the same challenge. In the MCP project’s July 2026 specification update, maintainers identify authorization as a substantial source of implementation effort and describe several changes to harden the flow. This is a protocol-level example of why enterprises need an architecture that carries identity, delegation, and policy across systems.

A useful reference architecture gives teams a common starting point. It should describe security functions and the interfaces between them, without requiring one vendor’s product stack. That is why Flemming Andreasen and I developed the Identity-Centric Security Framework for the Agentic Enterprise: to connect agent execution, identity, authorization, policy, enforcement, connectivity, security intelligence, operational intelligence, and enterprise management as cooperating planes. The framework describes those planes as logical functions connected through open interfaces, creating room for third-party integration and incremental adoption.

Consider an agent asking an MCP server to perform a sensitive operation. The managed agent presents a verifiable workload identity. An authorization service can issue a short-lived, narrowly scoped access token for the requested resource, tied to the agent and, where relevant, the human or organization that delegated the task. Enforcement checks identity and evaluates contextual information to determine if the request should be allowed to proceed. The interaction is logged so teams can connect the action to the agent, owner, target, and policy decision. If an agent is unmanaged, it can still be discovered, observed, and restricted while it lacks a verifiable identity.

Implementation details will vary. Vendor-neutral reference models can help teams compare approaches, spot integration gaps, and share a vocabulary across identity, networking, application, and security teams. That is how we move from principles to systems that can be tested, audited, and improved.

Read the full technical whitepaper, Identity-Centric Security for the Agentic Enterprise, for the complete architecture and access flows.

── more in #ai-agents 4 stories · sorted by recency
── more on @cisco 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/building-a-security-…] indexed:0 read:3min 2026-11-17 · —