The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Documentation in reference/custom_tools.md contains an example Python tool (tools/calculator.py) that uses the eval() function to process input strings, which can be exploited for arbitrary code execution.
[EXTERNAL_DOWNLOADS]: The framework automatically installs Python dependencies based on PEP 723 inline metadata in tool scripts and demonstrates fetching/running remote MCP packages using npx.
[COMMAND_EXECUTION]: The framework is designed to execute arbitrary shell commands for task verification and tool implementation as defined in field.toml configurations.
[DATA_EXFILTRATION]: As stated in reference/trajectory_analysis.md, the system records all agent interactions into local trajectory files in ~/.portlang/trajectories/, which may include sensitive data like API keys and credentials.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface where untrusted data enters the agent context via the re_observation primitive. The framework lacks explicit boundary markers or instructions to ignore embedded commands in these observations, and the agent possesses significant capabilities including shell command execution.

portlang