Category 1: Instructional Manipulation (Prompt Injection)
The most well-known agentic risk. Because agents process external data (like web pages or emails) to complete tasks, an attacker can embed malicious instructions in that data. If the agent follows the injected instructions, it can be coerced into using its tools to harm the organization.
Category 2: Tool Call Abuse
Agents are only as dangerous as the tools they have access to. Risk occurs when agents are over-provisioned with permissions—such as an HR agent with "admin" access to the employee database. An autonomous system making reasoning-driven tool calls requires a least-privilege architecture.
Category 3: Closed-Loop Instability
Agents often operate in a loop: plan, act, observe, repeat. If an agent encounters an error or ambiguous feedback, it may enter an unstable state, retrying actions rapidly or escalating its behavior in an attempt to reach the goal. This can lead to resource exhaustion, data corruption, or "hallucinated" system updates.
Category 4: The Accountability Gap
When an autonomous agent takes an action, identifying why it did so is difficult. LLMs are non-deterministic. Without a specialized audit trail like the one provided by Shield Control, organizations cannot provide the evidence required by regulators for automated decisions with significant legal effects.
