The Rise of Agentic Identity Hijacking: A Runtime Trust Framework for Securing Autonomous AI Agents
Year: 2026
Type: conference-paper
Abstract: Autonomous AI agents are increasingly used as enterprise identities with access to OAuth tokens, API keys, and delegated permissions across cloud and SaaS environments. However, existing Identity and Access Management (IAM), Privileged Access Management (PAM) frameworks authenticate agents only at session initiation, and even Zero Trust's continuous per-request checks do not verify whether a sequence of runtime actions remains aligned with authorized intent. This creates a new security threat called Agentic Identity Hijacking (AIH), where a legitimate AI agent identity is manipulated to perform unauthorized actions such as privilege escalation, lateral movement, reconnaissance, or task deviation while remaining within trusted execution paths. Recent incidents, including the 2025 GitHub Copilot vulnerability (CVE-2025-53773), demonstrate the risks posed by compromised autonomous AI behavior. To address this challenge, this paper proposes the AI Identity Runtime Trust Framework (AIRTF), a runtime security architecture integrating directed identity graphs, Intent Alignment Vector scoring, Bayesian Dynamic Trust Scoring, and per-action policy enforcement. AIRTF continuously evaluates agent behavior using intent alignment, credential integrity, behavioral drift, and escalation signals to make real-time trust decisions. Evaluated on the synthetic AIDE-1 enterprise dataset, AIRTF achieved 96.8% detection accuracy, an F1-score of 0.971, a false positive rate of 1.7%, and an average enforcement latency of 4.3 ms, outperforming traditional IAM, PAM, UEBA, and Zero Trust approaches.
Language: English
Topic: Access Control and Trust
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7226036498
