The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by processing untrusted binary files provided by users.
Ingestion points: Files binary_ops.py (import command), structural.py (metadata and structural queries), and functions.py (decompilation and disassembly) read and parse content from user-supplied binary samples.
Boundary markers: The skill architecture enforces a deterministic boundary by using structured JSON envelopes for all CLI tool outputs. The SKILL.md instructions explicitly command the agent to distinguish between 'CLI Evidence' and 'Agent Assessments' and to avoid inventing facts.
Capability inventory: The skill possesses capabilities that could be abused if an injection were successful, including arbitrary file writes within workspace boundaries (atomic.py), local network operations for toolchain setup (bootstrap.py), and execution of system commands (bootstrap.py, worker.py).
Sanitization: Content from binaries is typically hex-encoded or Base64-encoded for raw data, and strings are extracted and returned as escaped text. The path_security.py module prevents the skill from reading files in system-sensitive locations.
[COMMAND_EXECUTION]: The skill uses system commands for legitimate setup and process management tasks.
Evidence: scripts/binary_analysis/cli/bootstrap.py executes pip install to install the pyghidra bridge, and scripts/binary_analysis/bootstrap/deps.py executes commands like java -version to detect dependencies.
Context: These commands are part of the intended setup workflow and use list-based arguments to prevent shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The setup workflow references external repositories for required software components.
Evidence: scripts/binary_analysis/cli/bootstrap.py refers to official sources like adoptium.net (for Java JDK) and ghidra-sre.org (for the Ghidra framework).
Context: These references target established organizations and are used only for intended toolchain installation guidance.

binary-analysis