The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted source code and security reports that could contain malicious instructions. It implements defensive 'content envelopes' to ensure the agent treats all target material as data rather than instructions, mitigating potential injection attacks from the code being analyzed.
[COMMAND_EXECUTION]: The skill uses a set of specialized utility scripts located in the libexec/ directory and executes inline Python code for data processing and tool integration. These are documented as part of the core security analysis pipeline and are used to interface with the raptor framework.
[DATA_EXPOSURE]: The skill includes functionality for runtime observation of target binaries. This feature is designed for security research and is explicitly described as an opt-in process requiring operator consent to execute code within a sandbox environment.

code-understanding