The Agent Skills Directory

[PROMPT_INJECTION]: The documentation file reference/ai-red-teaming.md includes strings commonly used in jailbreak attempts, such as 'Ignore all previous instructions' and 'DAN'. These are explicitly labeled as adversarial test cases for use in AI security evaluations and do not target the agent's internal control logic.
[COMMAND_EXECUTION]: The skill provides detailed descriptions of command execution and lateral movement techniques based on the MITRE ATT&CK framework. These descriptions are intended to assist in designing attack scenarios and do not contain scripts or payloads for automated execution.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an ingestion point for system descriptions and requirements to generate threat models, creating a potential surface for indirect injection. However, the instructions mandate the use of structured frameworks (STRIDE, PASTA, ATLAS) which provides a conceptual layer of validation against malicious input data.

breach