Autonomous Security Agents in Cloud Environments: Leveraging Machine Learning and Agentic AI

Abstract

Autonomous security agents in cloud environments can be organized as a three-layer pipeline in which machine-learning-based detection identifies suspicious activity, agentic AI selects investigative steps and proposes responses, and a separately governed control layer authorizes and verifies consequential actions. This study specifies that integration boundary and evaluates the response-control layer using a narrow AWS IAM role-session operation. The ML interface carries detector version, score semantics, feature provenance, and supporting events; the agentic interface carries typed evidence requests and action proposals but no execution credentials. The control contract binds admitted evidence, current resource context, exact policy generation, action-specific approval, intervention capacity, worker fencing, and uncertain-result reconciliation. Evaluation includes 32 predicate fixtures, all 2,187 assignments of seven three-valued conditions, 66 malformed-input checks, 29 composed schedules, and structural and source-level policy mutations. The full transition configurations contain 19,516 reachable states and 76,948 transitions; five specified invariants hold within those bounds. In a separate 12-schedule comparison, admission-snapshot authorization produces four invalid issues and seven invalid effects, whereas the proposed issue-time controller produces no invalid issues but retains three post-issue invalid effects; an idealized atomic effect-boundary reference prevents those effects under a stronger modeled capability. All configurations preserve the three authority-preserving completions. The comparison extends the base paper's distinction between endpoint success and authorized completion to a concrete IAM contract; it neither reproduces that paper's agent experiments nor establishes superiority over its monitor. No cloud service, trained detector, language-model agent, or human-participant experiment was executed. The results therefore validate bounded control-layer behavior and its interfaces, not gains in detection accuracy or agent reasoning. The framework shows how ML findings and agentic investigation can be leveraged for autonomous cloud security without treating model outputs as execution authority.