The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted binaries, extracting strings and decompiled code for the agent to review. This data may contain hidden instructions or adversarial content meant to manipulate the agent's analysis.
Ingestion points: Files resources/scripts/strings-analyzer.py, resources/scripts/ghidra-headless.sh, resources/scripts/radare2-triage.py, and resources/scripts/ioc-extractor.js ingest binary data for processing.
Boundary markers: Absent. The skill does not provide specific instructions to help the agent distinguish between data and embedded commands in the output.
Capability inventory: The skill uses tools like Bash, Task, and filesystem which allow for significant interaction with the host environment.
Sanitization: Absent. Untrusted content from binaries is not sanitized before being presented to the agent.
[DYNAMIC_EXECUTION]: The resources/scripts/ghidra-headless.sh script generates temporary Python scripts (decompile_all.py, export_callgraph.py) from hardcoded templates to automate Ghidra tasks. These generated scripts are then executed at runtime.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to run local scripts that invoke external analysis binaries like ghidra, radare2, and strings.
[EXTERNAL_DOWNLOADS]: The skill's documentation identifies dependencies on standard libraries and tools, recommending installation via pip and npm from public registries for packages such as r2pipe, pefile, and axios.
[METADATA_POISONING]: The author field in SKILL-meta.yaml identifies 'ruv' as the author, which differs from the 'dnyoussef' attribution provided in the platform context.

reverse-engineering-quick-triage