The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides various tools for executing arbitrary shell commands and subprocesses within cloud containers. Evidence: references/images.md shows the use of .run_commands() during container image building, and references/web-endpoints.md includes examples using subprocess.Popen() to launch arbitrary processes.
[CREDENTIALS_UNSAFE]: Reference documentation (references/images.md) contains specific instructions that encourage users to upload sensitive local directories, such as the AWS credentials folder ~/.aws, to the cloud environment, posing a high risk of credential exposure.
[REMOTE_CODE_EXECUTION]: The core functionality of the skill is designed to facilitate the execution of arbitrary Python code and shell scripts on remote serverless infrastructure.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading external code and resources, including the installation of the modal package and cloning external repositories from GitHub during the build process.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection (Category 8).
Ingestion points: Data is ingested through user-defined web endpoints in references/web-endpoints.md (e.g., predict(data: dict)).
Boundary markers: None are present in the provided examples to distinguish untrusted data from processing instructions.
Capability inventory: The skill possesses extensive capabilities, including file system access, network operations, and command execution across its various reference scripts.
Sanitization: There is no evidence of input validation or escaping before external data is processed in the provided code snippets.

modal