The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to download the sasquatch tool and binwalk installation instructions from repositories hosted on GitHub by non-trusted entities (devttys0 and ReFirmLabs).
[REMOTE_CODE_EXECUTION]: Instructions include cloning a remote repository and immediately executing a build script (./build.sh), which runs external code on the local system.
[COMMAND_EXECUTION]: The skill uses various shell commands and scripts to perform binary analysis, carving, and filesystem extraction, which could be exploited if malicious inputs are provided.
[PRIVILEGE_ESCALATION]: The skill requires the use of sudo for administrative tasks such as installing system packages (apt-get) and mounting filesystem images, providing a path to high-level system access.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted firmware binaries that may contain hidden instructions or malicious structures designed to manipulate the agent's output or exploit the parsing tools.
Ingestion points: Untrusted firmware files (e.g., firmware.bin) are loaded for entropy analysis and signature scanning.
Boundary markers: There are no specific instructions or delimiters used to prevent the agent from following natural language instructions that might be embedded in the binary data (e.g., in strings or metadata).
Capability inventory: The agent has access to shell execution, filesystem mounting, network downloads via git/apt, and Python script execution.
Sanitization: The process relies on external tools like binwalk and unsquashfs to parse data; it does not implement additional validation or sanitization of the binary content before processing.

firmware-extraction