04-reverse-engineering
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it is designed to ingest and analyze untrusted data from binary artifacts, firmware, and protocol captures. \n
- Ingestion points: The skill reads external files through
scripts/binary_analyzer.pyand standard utilities likestrings,binwalk, andobjdump. \n - Boundary markers: The skill includes an "Evidence contract" section that explicitly instructs the agent to treat tool outputs and file contents as untrusted evidence, separating observed data from instructions. \n
- Capability inventory: The skill performs file system reads, string extraction, and binary parsing. It also recommends the use of external debuggers and frameworks such as GDB, Ghidra, and radare2. \n
- Sanitization: The skill does not perform automated sanitization of extracted strings but provides procedural instructions to the agent to prevent instruction override. \n- [EXTERNAL_DOWNLOADS]: The documentation recommends installing established, well-known third-party Python libraries including
capstone,pyelftools,pefile, andlief. These are industry-standard dependencies for binary analysis and disassembly.
Audit Metadata