The Agent Skills Directory

[COMMAND_EXECUTION]: The skill incorporates helper scripts scripts/go_binary_triage.py and scripts/macos_app_triage.py which utilize the Python subprocess module to run diagnostic utilities. Evidence shows the invocation of tools including file, shasum, go, strings, otool, readelf, hdiutil, plutil, codesign, and spctl. These commands are strictly scoped to static analysis tasks and are designed to inspect the target files without executing them.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for processing untrusted binary data and metadata which could theoretically contain instructions targeting the LLM. It mitigates this risk by establishing clear boundary markers, such as instructions to avoid printing raw strings to the conversation and instead saving outputs to a workspace for specific querying. Capability inventory is restricted to file system operations in /tmp and execution of localized static tools. Sanitization is achieved by the agent acting as a filter, using tools like rg to extract specific leads rather than ingesting raw, potentially malicious byte streams.

reverse-engineering