The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted firmware images and tool reports, creating a vulnerability surface where malicious metadata in firmware could influence agent behavior.
Ingestion points: Firmware images processed in SKILL.md and report.json parsed in scripts/summarize_unblob.py.
Boundary markers: SKILL.md explicitly instructs the agent to treat file names and tool messages as input data, not instructions.
Capability inventory: The skill can execute shell commands (unblob, binwalk), write files, and use sudo for mounting filesystems.
Sanitization: The summarize_unblob.py script implements truncation for long strings and limits the number of items processed to reduce injection risks.
[COMMAND_EXECUTION]: The skill utilizes several shell commands for firmware extraction and metadata calculation.
Evidence: Commands such as unblob, sha256sum, binwalk, and dd are central to the skill's operations in SKILL.md and references/manual-extraction.md.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing third-party security tools from well-known repositories.
Evidence: Instructions for pipx install 'unblob==26.6.4' and git clone https://github.com/devttys0/sasquatch are provided in references/unblob.md and references/manual-extraction.md.
[PRIVILEGE_ESCALATION]: Manual extraction procedures include necessary administrative commands for tool installation and filesystem access.
Evidence: Use of sudo apt-get install and sudo mount in references/manual-extraction.md.

firmware-extraction