Cisco Ships Build-Time Policy Enforcement for Agent Frameworks, Extending the Governance Stack

Cisco announced at RSAC 2026 an Agent Runtime SDK that embeds policy enforcement into agent workflows at build time, supporting AWS Bedrock AgentCore, Google Vertex Agent Builder, Azure AI Foundry, LangChain, and the Model Context Protocol (MCP). Cisco also released DefenseClaw, an Apache-2.0 open-source runtime threat-detection framework with 843 stars and 151 forks that enforces block/allow lists in about two seconds without an agent restart, and the AI Defense Explorer Edition, which runs self-service red teaming across more than 200 risk subcategories in about 20 minutes. The announcements target a gap Cisco's survey identified: 85% of enterprises are experimenting with AI agents but only 5% have moved them into production.

Eighty-five percent of enterprises are currently experimenting with AI agents, yet only five percent have successfully moved those agents into production Cisco survey . This 80-point chasm is often misdiagnosed as a technical hurdle, but the reality is a profound trust deficit. Organizations are not struggling to build agents; they are struggling to govern them at scale. Cisco’s recent announcements at RSAC 2026, including the Agent Runtime SDK https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m03/cisco-reimagines-security-for-the-agentic-workforce.html , suggest a shift in how we approach this gap. By embedding policy enforcement directly into agent workflows at build time — supporting frameworks like AWS Bedrock AgentCore, Google Vertex Agent Builder, Azure AI Foundry, LangChain, and the Model Context Protocol MCP https://forkast.news/glossary/model-context-protocol-mcp/ — the industry is beginning to treat governance as a distinct infrastructure layer. This is not merely a feature update; it is the formalization of a necessary stage in the governance stack https://forkast.news/the-agent-governance-stack-is-forming-four-products-two-weeks-one-pattern/ . The emerging stack now follows a clear progression: governance specifications like OWASP and NIST define the rules, while build-time enforcement, such as the Agent Runtime SDK, bakes those rules into the agent’s DNA. From there, the stack moves to runtime authority via tools like Akeyless, runtime observation and enforcement through platforms like Akamai and CrowdStrike, and finally, a control plane provided by WSO2 https://forkast.news/wso2-ships-an-open-source-control-plane-formalizing-agent-governance-as-infrastructure/ . By integrating build-time enforcement, Cisco is filling a critical void between abstract policy and active execution. The deeper issue is that build-time enforcement, while necessary, is not sufficient. Even a perfectly governed agent at the moment of deployment is subject to runtime drift, where an agent’s behavior evolves beyond its initial security parameters due to changing environmental inputs or model updates. This drift can lead to unauthorized tool usage, data exfiltration, or unexpected cross-agent interactions that were never validated during the build phase. Cisco addresses this with DefenseClaw https://github.com/cisco-ai-defense/defenseclaw , an open-source Apache-2.0 framework that currently boasts 843 stars and 151 forks. DefenseClaw provides comprehensive runtime threat detection through a suite of specialized engines. It includes a Skill Scanner to monitor agent capabilities, an MCP Scanner to check servers for malicious actions, an A2A Scanner for agent-to-agent communication, CodeGuard for static analysis of AI-generated code, and an AI BoM generator for tracking models, memory, and dependencies. The framework supports a wide range of agents, including Claude Code, OpenAI Codex, Cursor, Windsurf, Gemini CLI, GitHub Copilot, Hermes, OpenClaw, and Zeptoclaw. It enforces block/allow lists in approximately two seconds without requiring an agent restart, while also hooking into NVIDIA OpenShell for automated security and exporting telemetry to Splunk. This push toward accessible security tools is lowering the barrier for builders. The AI Defense Explorer Edition https://blogs.cisco.com/ai/introducing-cisco-ai-defense-explorer provides self-service, algorithmic red teaming across more than 200 risk subcategories in about 20 minutes. By offering the same capabilities as the enterprise edition with CI/CD integration for GitHub Actions, GitLab, and Jenkins, Cisco is enabling developers to stress-test their agents before they ever reach a production environment. This infrastructure evolution is essential because, as Jeetu Patel noted https://blogs.cisco.com/news/reimagining-security-for-the-agentic-workforce , AI agents represent a new workforce of co-workers that dramatically expand organizational capacity. However, that expansion is only sustainable if the underlying infrastructure can handle the complexity of agent governance https://forkast.news/glossary/agent-governance/ . While tools like Temporal https://forkast.news/temporal-raises-550m-at-12-55b-valuation-signaling-durable-execution-as-core-agent-infrastructure/ provide the necessary runtime infrastructure for reliable workflows, the security layer must remain distinct to prevent policy from becoming a bottleneck. The 85%/5% gap will not close through better models alone. It will close when builders can rely on a modular, layered governance stack that enforces policy at build time and monitors for drift at runtime. By separating these concerns, the industry is finally moving toward a model where trust is an architectural component rather than an afterthought.