The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze untrusted binary files (libapp.so, kernel_blob.bin) from third-party applications. Maliciously crafted strings or metadata embedded in these files could theoretically attempt to influence the agent's behavior if they contain prompt-like instructions that the LLM processes during analysis.
Ingestion points: The agent reads data from libapp.so and kernel_blob.bin using objcopy, strings, and python3 scripts.
Boundary markers: The instructions do not specify any delimiters or safety warnings to wrap content extracted from the binary files to prevent the agent from following instructions found within them.
Capability inventory: The skill utilizes file system access, subprocess execution of binary utilities, and local Python script execution.
Sanitization: There is no evidence of sanitization or filtering for the strings extracted from analyzed binaries before they are presented to the agent.
[COMMAND_EXECUTION]: The skill employs standard, legitimate system utilities for binary analysis, including readelf, objcopy, strings, unzip, and file. These tools are used for their intended purposes in a reverse engineering workflow.
[DYNAMIC_EXECUTION]: The skill contains hardcoded Python scripts executed via heredocs (`python3
<<'PY'`). These scripts are used to parse specific binary structures by scanning for magic numbers and calculating offsets, which is a standard and safe technique for this domain.

re-flutter