The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from SVG and PNG assets, creating a potential surface for indirect prompt injection where malicious metadata or file structures could be designed to influence the agent's logic.
Ingestion points: External files imported via the --asset flag and JSON-based workflow plans provided in the examples/ directory.
Boundary markers: The system utilizes manifest.json for mandatory digest binding and requires explicit integer-based object ID selection (select.set) to prevent instruction leakage from object properties.
Capability inventory: The skill can execute a native CLI for rendering, perform filesystem writes for export formats (SVG, PNG, PDF), and read registered assets via scripts/commands.py and scripts/workflow.py.
Sanitization: All plans and assets undergo SHA-256 integrity checks, and the skill performs strict validation of native command parameters before execution.
[SAFE]: The skill enforces security hardening by requiring python3 -I -B for its internal scripts. The -I (isolated) flag prevents the agent from being influenced by environment-based Python tampering, while -B prevents the creation of potentially exploitable bytecode cache files.
[SAFE]: The bootstrap process for the native CLI implements a "locked" release mechanism. It verifies the SHA-256 digests of all downloaded binaries against a pinned runtime.lock.json file before execution, mitigating supply chain risks.

vectorcraft-use