A practical series examining how vulnerabilities in AI-enabled and agentic applications can expose data, bypass access controls, and enable unauthorized actions.
Each breakdown focuses on a specific security weakness and examines how it can be reproduced in a controlled environment.
The analysis connects vulnerable behavior to its underlying cause, then examines the hardened implementation and tests whether the fix prevents the unauthorized behavior.
The practical scenarios are developed in AgenticEnclave, a security lab focused on AI-enabled and agentic applications. The source code provides access to the lab's implementations.
Each scenario examines a distinct security weakness in AI-enabled or agentic applications. Available scenarios include practical demonstrations; upcoming scenarios will be added as they are implemented.
Examine how missing resource-ownership checks can allow one customer to cancel another customer's order, and how authorization enforcement prevents the unauthorized action.
Examine how inadequate authorization and excessive permissions can allow operations beyond an application's intended access boundaries.
Examine how weaknesses in AI-enabled applications can expose sensitive information that should remain protected.
Explore how missing workflow constraints and incomplete business-rule enforcement can enable operations that should not be permitted.
Examine how malicious instructions supplied directly to an AI system can influence its behavior and challenge intended controls.
Explore how malicious instructions embedded in external content can influence an AI system when that content enters its context.
Examine how excessive autonomy, permissions, or execution capabilities can allow an agent to take actions beyond what a task requires.
Examine how inadequate controls around tool selection and execution can allow an agent to invoke operations it should not perform.
Explore how weaknesses in trust boundaries and authorization between agents can enable access to capabilities or resources beyond an agent's intended privileges.
Examine the risks of using AI-generated output without appropriate validation, sanitization, or security controls at downstream execution boundaries.
The techniques demonstrated in this series are intended for authorized testing and controlled environments. Never test systems you do not own or have explicit permission to assess.