The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Python scripts (scripts/macos_app_triage.py and scripts/go_binary_triage.py) to perform automated binary analysis. These scripts execute a variety of system utilities (e.g., otool, nm, hdiutil, strings, go) via subprocess.run.
The scripts employ list-based command construction and path validation to mitigate shell injection risks.
Commands are restricted to static analysis tasks, and the scripts explicitly avoid launching the target binaries.
The macOS triage script includes a controlled disk image mounting procedure using hdiutil with safety flags (-readonly, -nobrowse, -noautoopen) and ensures mandatory cleanup of the mountpoint.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted binary data, creating a potential surface for indirect prompt injection via strings or metadata extracted during triage.
Ingestion points: Metadata and strings are extracted from user-supplied binaries using the triage scripts and standard tools like strings.
Boundary markers: SKILL.md contains rigorous guidelines for the agent, including explicit 'Evidence Levels' and a 'Default Posture' that mandates treating 'do not run this binary' as a hard constraint.
Capability inventory: The skill can read local files, create temporary workspaces, and execute local triage tools. It does not possess network exfiltration or arbitrary code execution capabilities.
Sanitization: Instructions mandate the redaction of credential-like values and recommend using rg to filter output, reducing the risk of the agent incorrectly interpreting malicious strings as instructions.

reverse-engineering