The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The primary function of this skill is to ingest and process untrusted binary data (from IDA Pro databases) for consumption by an AI. This creates a known attack surface where malicious strings or code patterns within a binary could be used to influence an AI's behavior during analysis.
Ingestion points: The skill reads function names, decompiled C code, disassembly, and strings from IDA's database, which are derived from untrusted binaries.
Boundary markers: While the export format includes metadata headers (e.g., func-name, callers), it does not include explicit delimiters or instructions to the AI to ignore instructions potentially embedded within the data itself.
Capability inventory: The skill writes various files (C code, ASM, memory dumps, logs) to the local filesystem.
Sanitization: The implementation includes filename sanitization to prevent directory traversal or invalid file paths, but it does not perform content sanitization to strip potential prompt injection sequences from the exported text.
[METADATA_POISONING]: There is an inconsistency between the skill's description, which claims it is a "Skill by ara.so", and the reported author of the skill. Such discrepancies can be misleading regarding the provenance and trust level of the tool.
[DYNAMIC_EXECUTION]: The documentation encourages the use of execfile() to run the exporter script within IDA Pro. This pattern involves dynamically executing code from a file path, which is a common but risky practice if the target file is not properly secured.

ida-no-mcp-decompiler-exporter