The Agent Skills Directory

Indirect Prompt Injection (SAFE): The skill identifies an attack surface by processing untrusted binary data, which is inherent to its primary purpose.
Ingestion points: Untrusted content is ingested via get-strings, get-symbols, and get-decompilation tools.
Boundary markers: While the workflow is structured, it does not explicitly use delimiters or instructions to prevent the agent from obeying strings found inside the binary.
Capability inventory: The skill possesses the ability to read binary metadata and write a prioritized task list using TodoWrite.
Sanitization: No explicit sanitization or filtering of the binary data (e.g., escaping control characters in strings) is performed before presentation.
Conclusion: As this is a specialized security triage tool, the processing of untrusted data is the intended function. The risk of the agent being influenced by instructions embedded in the binary is a known trade-off of LLM-based reverse engineering.

binary-triage