CrowdStrike launches identity controls for AI agents

CrowdStrike has launched Continuous Identity for AI Agents, extending its identity security controls to AI agents.

The offering sits within CrowdStrike's Falcon Next-Gen Identity Security range and is designed to govern how AI agents access systems, data and applications across on-premises, software-as-a-service, browser and cloud environments.

At the centre of the product is a model of continuous authorisation rather than one-time approval. The system evaluates each action by an AI agent in real time based on who owns the agent, who is calling it and the risk posture of the device involved.

The launch reflects a broader shift in cyber security as companies deploy software agents that can carry out tasks, invoke tools, call application programming interfaces and pass work to sub-agents with limited human intervention. That has created a challenge for access control systems built around static credentials and fixed permissions.

Access model

The system can grant, deny or revoke access dynamically as risk changes. It is also designed to remove standing privileges, a long-standing security issue in which users or services retain access rights they do not need at a given moment.

According to CrowdStrike, each AI agent is assigned a verifiable identity based on the SPIFFE standard, which is used to issue and manage workload identities instead of relying on static credentials such as API keys. The authorisation context remains attached when an agent delegates work to a sub-agent.

The product also ties into Falcon AI Detection and Response, which monitors prompts and intent for signs of permission misuse or attempts to push a large language model beyond its approved role. If suspicious behaviour is detected, the system can revoke access.

CrowdStrike linked the launch to technology gained through its acquisition of SGNL. The deal gave the company technology for risk-aware access decisions that can be applied across human users, machine identities and now AI agents.

Growing focus

Identity has become a more prominent battleground in cyber security as attacks increasingly target logins, credentials and privileged access rather than only endpoints. The rise of AI agents adds another layer because these systems can act quickly and reach broadly across business tools.

That creates a different risk profile from conventional software automation. An AI agent may not only retrieve information, but also make decisions, trigger actions in connected systems and hand tasks to other agents, complicating the chain of responsibility when something goes wrong.

CrowdStrike is positioning the product as a control layer for those agent-driven workflows. Rather than assuming an approved identity remains safe after initial login, its model checks trust continuously during access, privilege escalation and lateral movement.

Elia Zaitsev, Chief Technology Officer at CrowdStrike, outlined the company's view of that shift in identity management.

"AI agents are transforming how work gets done, and how identities must be secured," Zaitsev said. "Point-in-time authorisation becomes a legacy approach the second agents are given autonomy. Authorise once and trust indefinitely is not a security model, it's a liability. That's the shift CrowdStrike is driving, from static, one-time access decisions to Continuous Identity."

The launch adds to a growing contest among cyber security providers to adapt identity and access management tools for AI systems. Vendors across the sector are trying to address how autonomous and semi-autonomous software should be authenticated, monitored and limited as use spreads across enterprise technology estates.

For CrowdStrike, the announcement also broadens Falcon's role beyond endpoint and cloud workload security into identity-based control over automated actors. Continuous Identity for AI Agents applies the same risk-aware authorisation approach across human, non-human and AI identities, covering the path from initial access to privilege escalation and lateral movement.