Beyond human insiders: Emerging autonomous AI risks to organisations

Autonomous artificial intelligence (AI) agents are increasingly embedded in organisational workflows, performing complex tasks such as software development, data analysis, cybersecurity monitoring, and process automation. Unlike traditional software tools, these systems operate with a high degree of autonomy, executing multi-step reasoning and interacting directly with organisational systems using legitimate access privileges. While these capabilities enhance efficiency and enable new forms of digital work, they also introduce a new class of security risks that extend beyond traditional threat models.

Recent incidents demonstrate that AI agents with authorised access can produce harmful outcomes without external intrusion, challenging conventional insider threat frameworks. For example, in the Google Gemini memory-poisoning case (2025), manipulated contextual inputs influenced the agent’s reasoning, resulting in unintended and potentially harmful outputs. Similarly, the Devin AI "terminal takeover" case (2025) illustrates how an autonomous coding agent executed system-level commands beyond intended boundaries, effectively bypassing operational safeguards.

These cases highlight a critical shift: insider-like risks are no longer limited to human actors. Instead, autonomous AI systems can generate harmful outcomes through their internal reasoning processes and interactions with complex environments. Errors in interpretation, optimisation, or contextual understanding may lead to unintended actions, even when the system operates with legitimate access. This exposes a significant gap in organisational security, as existing insider threat models are primarily designed around human intent and behaviour.

To address this gap, this study introduces the concept of the Digital Insider, defined as an autonomous AI agent capable of generating insider-like risks within organisational systems. Using rigorous qualitative analysis grounded in the Gioia methodology, the study systematically analysed documented incidents from the AI Incident Database to identify three key mechanisms through which these risks emerge:

• Reasoning Deviation - where AI agents diverge from intended objectives due to misinterpretation, flawed optimisation, or hallucinated system states
• External Exploitation - where malicious actors deliberately leverage AI agents to perform harmful activities, automating cyber espionage or attack execution
• System-Level Coercion – where compromised infrastructure, malware, or unsafe configurations influence and alter AI agent behaviour

Building on the CERT insider threat model, the proposed Digital Insider framework extends traditional theory by incorporating AI reasoning drivers while preserving the structure linking drivers, decision processes, and system actions. This enables a unified understanding of insider risk that includes both human and non-human actors.

The findings demonstrate that AI-driven insider-like incidents arise through the interaction between reasoning processes, autonomous decision-making, and socio-technical environments. These risks represent an emerging class of organisational threats as AI adoption accelerates across organisations worldwide.

By moving beyond human-centric models, this research highlights the urgent need for organisations to rethink insider threat management and develop governance approaches suited to autonomous AI systems. The Digital Insider framework provides both theoretical and practical foundations for organisations to classify, manage, and govern agentic AI within existing security and risk management structures.