What Cloud Security Actually Looks Like in 2026
Cloud security has been redefined by three structural shifts that happened fast enough to outpace most security programs' ability to adapt. Not three new vulnerability classes or three new tool categories, three changes to the fundamental operating model of cloud infrastructure that render significant portions of the conventional cloud security playbook insufficient. The first shift: identity re…
- Identity replaced network perimeter as the primary cloud security control plane, IAM policy quality is more determinative of security posture than network segmentation quality.
- Ephemeral infrastructure requires policy-based security rather than state-based security, you review the policy governing the class of workloads, not the state of individual running instances.
- AI expanded the attack surface across three vectors: AI agents with tool access, AI-generated code, and AI-generated infrastructure configuration, each requiring different security approaches.
- Continuous policy verification is the operating model that addresses all three structural shifts, real-time posture awareness rather than periodic snapshots.
- The organizational model for cloud security has shifted: platform teams own verification infrastructure, security teams own policy, and developers receive violations as operational feedback rather than audit findings.
Identity-Centric Security: What It Actually Means Operationally
Identity-centric security is a widely used term that often means different things to different teams. Operationally, it means that the primary mechanism for making security decisions, who can access what, under what conditions, with what constraints, is the identity and authorization layer rather than the network layer. In practice, this means that the quality of your IAM design is more determinative of your security posture than the quality of your network segmentation. A service that has netw
Ephemeral Infrastructure and What It Breaks
Ephemeral infrastructure breaks several assumptions that conventional cloud security relied on. The assumption that you can review an instance's configuration and assess its security posture doesn't hold when that instance might not exist tomorrow. The assumption that security monitoring can track a running workload's behavior over time doesn't hold when the workload might be a function that runs for 200 milliseconds. The assumption that access controls can be audited against running systems doe
AI as an Attack Surface Expansion
The AI attack surface in cloud environments has three distinct components that require different security approaches. The first component is AI agents with tool access. An AI agent that can call APIs, execute code, read and write data, and make external requests has a blast radius proportional to the permissions of its underlying service account or IAM role. The security model for AI agents is the same as for any other automated actor in the system: least privilege access, audit logging of all
Frequently asked questions
- How do you build an identity-centric security program if you're currently network-perimeter focused?
- The transition doesn't require dismantling network controls, it requires reordering the design priorities. Start by treating every IAM policy as a security control with the same review discipline as a firewall rule. Map the blast radius of each IAM role: if this role were compromised, what could an attacker do? The roles with the largest blast rad…
- What does least-privilege look like for ephemeral workloads like Lambda functions?
- For ephemeral workloads, least privilege means scoping the execution role to the specific API calls the function makes, to the specific resources it accesses, and for the specific conditions under which it runs. The IAM policy for a Lambda function should list every AWS API call the function makes, with resource ARNs scoped to the specific resourc…
- How should the cloud security program account for AI agents without knowing what agents will be deployed in the future?
- Design for the class rather than the instance. Instead of reviewing each AI agent deployment individually, define an organizational policy for what any AI agent deployed in your environment can do: what IAM permission level, what network access, what data access, what logging requirements. Apply the policy through platform controls that enforce it…
- What is the most important operational change for cloud security programs in 2026 that wasn't necessary in 2022?
- Explicit governance for AI-generated artifacts in the cloud estate. In 2022, all code, configuration, and infrastructure definition that made it to production passed through an engineer's intent. In 2026, a meaningful fraction of those artifacts were partially or wholly generated by AI tools, with the engineer's review being one step in a process …