As the AI agent ecosystem expands, industry moves to secure AI agents are accelerating as expected.

At RSAC 2026, a global security event that opened on Sunday in San Francisco, major security vendors also rolled out a wave of announcements targeting AI agents.

A range of products and technologies were unveiled to help companies reduce risks from the spread of AI agents, from governance to identity (ID).

Microsoft upgraded a broad swath of its security portfolio under a vision that AI agents should be treated not simply as assets to protect but as a core security layer. The update includes agent governance, identity protection, data security, cloud and endpoint defence, and AI-based security operations capabilities.

The company said that as AI agents play a bigger role inside companies, a central control system for agent governance, stronger identity protection, tighter data protection and autonomous security operations are needed. Microsoft Agent 365, the centrepiece of the update, supports IT, security and business teams in managing AI agents deployed across an organisation from a single place.

• Microsoft steps up AI agent security push with major update

Cisco also used the event to unveil a range of AI agent security products. It introduced zero-trust access controls targeting AI agents, a new version of its AI model red-teaming tool, and AI agents tailored for security operations centres (SOC).

Cisco's view is that AI agent security is not finished with protecting the agent itself. If the underlying AI model an agent runs on is vulnerable or compromised, the agent can be put at risk as well. With that in mind, Cisco released a new version of AI Defense called "AI Defense: Explorer Edition". AI Defense: Explorer Edition supports developers, application security teams and security researchers in running red-team tests directly on AI models and applications before deploying them to production environments.

CrowdStrike, a cloud-based endpoint company, also announced an upgrade to its Falcon cybersecurity platform, considering a situation in which AI agents are spreading across endpoints, software as a service (SaaS) and cloud environments in corporate settings.

According to the company, as AI agents access data and gain high privileges, existing security controls are difficult to handle from a governance standpoint. CrowdStrike therefore put forward an endpoint-centred strategy for AI security controls. Because AI actions are ultimately executed on devices, the company said endpoints are the most suitable point for behaviour monitoring, policy enforcement and threat blocking.

Among the new features, "EDR AI runtime protection" tracks commands, scripts, file activity and network connections that AI applications and agents run on systems, providing security teams with runtime visibility. It can trace suspicious behaviour back to the originating process and isolate the endpoint. Runtime refers to the moment an application or agent is actually executing, meaning it is in operation.

Google Cloud, which recently completed its acquisition of Wiz, also unveiled an AI agent-based integrated security platform strategy at the RSAC conference. Google Cloud is emphasising both technologies that protect AI agents and security that uses AI agents.

As part of AI agent security, Google Cloud added an "AI Protection" function to Security Command Center and strengthened its capabilities to detect threats targeting AI agents. It also updated Model Armor, a service for protecting AI models, and reinforced functions to respond to risks such as prompt injection, sensitive data leakage and tool manipulation.

With its integration with Wiz, Google Cloud is also accelerating its AI security solution strategy spanning multicloud infrastructure.

Based on its latest Gemini AI model, Google introduced an adaptive AI agent that carries out alert investigations, intelligence synthesis and response tasks in real time. The strategy is to reshape security operations centres (SOC) that rely on existing static playbooks into agent-centred operations.