The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data in the form of compiled binaries or firmware images and extracts embedded strings, symbol names, and decompiled code into text and JSON files. When the agent reviews these output files, malicious content embedded inside the binary could potentially trigger unintended actions if interpreted as instructions.
Ingestion points: Binary files are supplied to ghidra-analyze.sh and processed by Ghidra's analyzeHeadless tool, generating text artifacts like _summary.txt, _interesting.txt, and _decompiled.c.
Boundary markers: Absent. The instructions do not prescribe or place delimiters or guardrails around the text outputs when the agent views them.
Capability inventory: The skill possesses the ability to execute shell scripts and view file contents using the Bash and Read tools.
Sanitization: The Java export scripts contain an escapeJson method to maintain valid structural JSON data, but they do not filter or sanitize the underlying semantic text or natural language string values extracted from the binary.
[COMMAND_EXECUTION]: The skill executes external processes by launching Ghidra's headless analysis executable through shell wrapper scripts (ghidra-analyze.sh and find-ghidra.sh). Arguments are constructed dynamically and passed into an execution array to handle project management and analysis workflows.

ghidra-headless