The AI agent security market crystallized at RSAC 2026. In the span of a single conference week, four major vendors announced distinct products aimed at the same problem: AI agents operating in enterprise environments with no identity, no accountability, and no visibility. What emerged wasn’t a consensus. It was a competition between four different theories about where AI agent security actually has to happen.
The four launches break down by architectural layer. Cisco introduced DefenseClaw, an open-source framework that sandboxes each agent at the infrastructure level and automates security inventory for agentic deployments. It’s built on NVIDIA OpenShell and enforces Zero Trust Access controls at the foundation. Per Cisco’s announcement, the framework establishes trusted identities for agents, solving the problem before agents ever touch sensitive data. Rubrik went a layer up. Its Semantic AI Governance Engine, SAGE, powers the new Rubrik Agent Cloud platform and replaces what Rubrik describes as “static, manual oversight” with intent-driven governance, the agent’s behavior is evaluated against semantic policy, not just access rules. Google Security moved to the detection and response layer, introducing new agentic SOC agents at RSAC 2026 designed to help security defenders focus on critical threats using frontline threat intelligence. Reco took a different approach entirely: SaaS visibility. According to press coverage, Reco’s AI Agent Security gives security teams monitoring and control over AI agents, including Microsoft Copilot and ChatGPT, operating across enterprise SaaS platforms.
Why this matters: enterprise security teams don’t yet have a standard for AI agent governance. These four launches represent the market’s first concentrated attempt to offer one, but they’re solving different parts of the problem. Cisco secures the deployment layer. Rubrik governs agent intent. Google automates the defender’s response to agent threats. Reco monitors what agents are already doing in your SaaS stack. None of them, alone, covers the full attack surface.
The timing is not accidental. Agentic AI deployments are moving faster than security frameworks. This pipeline has covered the pattern across multiple cycles – from Dapr Agents v1.0 reaching general availability, to Zuckerberg’s public case for personal AI agents, to enterprise-level discussions of agent governance. RSAC 2026 is the security industry catching up to what engineering teams started deploying months ago.
Watch for two things. First, whether any of these frameworks attempts interoperability, right now, a security team running Cisco DefenseClaw and Rubrik SAGE simultaneously faces a coordination problem those vendors haven’t solved publicly. Second, whether NIST or an industry working group produces agent identity and authorization standards that vendors can build to. Until that happens, the architecture layer you prioritize reflects a bet on which vendor’s theory of the problem turns out to be right.
The concentration of launches at RSAC 2026 signals that enterprise demand for agent visibility and control has reached the point where the market is moving. What it hasn’t yet produced is clarity, and that’s the real story for security architects evaluating these tools today.