The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents functions idapython_snippet and idapython_file that allow the execution of arbitrary Python code and scripts. This provides a mechanism for arbitrary command execution on the host machine if used by the agent to perform actions based on untrusted input.
[DATA_EXFILTRATION]: The function load_file_bytes is documented to allow reading arbitrary bytes from any host file path into memory, which can lead to the exposure of sensitive system files.
[DATA_EXFILTRATION]: Functions such as gen_listing and gen_cfg_dot_file allow the writing of data to arbitrary paths on the host filesystem, posing a risk of unauthorized file creation or data leakage.
[PROMPT_INJECTION]: The documentation describes several functions that read data from external binaries, creating a vulnerability surface for indirect prompt injection attacks where malicious data in a binary could influence agent behavior.
Ingestion points: SKILL.md (via functions like decompile, comment_at, grep, bytes, and name_at which retrieve content from binary files).
Boundary markers: Absent (there are no instructions for the agent to use delimiters or ignore instructions embedded within binary data).
Capability inventory: SKILL.md (the tool environment includes powerful write and execution capabilities such as idapython_snippet, idapython_file, load_file_bytes, and patch_byte).
Sanitization: Absent (no sanitization or validation logic is defined for the data extracted from binary files before it is processed by the agent).

functions