The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill documentation in SKILL.md includes the command sudo make install, which executes the installation process with administrative privileges to modify system directories.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and processing of raw text for tokenization and model training, creating a potential vector for indirect prompt injection.
Ingestion points: Raw text inputs are processed via sp.encode() and training corpus files (e.g., data.txt, corpus.txt) are used in spm.SentencePieceTrainer.train as shown in SKILL.md and references/training.md.
Boundary markers: The instructions do not define delimiters or special instructions to prevent the model from obeying instructions embedded in the processed text.
Capability inventory: The skill documentation covers local file system access for reading training data and command-line execution of the spm_train utility.
Sanitization: The examples do not demonstrate sanitization or filtering of input text before tokenization.
[EXTERNAL_DOWNLOADS]: The skill documentation contains commands to download source code from Google's official GitHub repository and install packages using pip.
[COMMAND_EXECUTION]: The documentation outlines the use of shell commands for building from source (cmake, make) and running training utilities (spm_train).

sentencepiece