The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses pickle.load() to deserialize a machine learning model from the local file path models/lgbm_quantile_baseline.pkl. This is an unsafe deserialization pattern that can lead to arbitrary code execution if a user or attacker replaces the model file with a malicious payload.
[EXTERNAL_DOWNLOADS]: The installation process involves cloning a remote repository from https://github.com/OranAi-Ltd/oransim.git and installing it in editable mode. This pulls the entire logic of the simulation engine from an external source.
[INDIRECT_PROMPT_INJECTION]: The skill employs "soul personas," which are LLM-backed agents designed to read and react to marketing creatives, establishing a vulnerability surface where instructions embedded in creative content could influence agent behavior.
Ingestion points: Creative data and metadata processed by the AgentSociety class and simulation endpoints (/api/predict).
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are documented for the consumer agent simulation.
Capability inventory: The system performs network requests to LLM providers (OpenAI, Anthropic, etc.) and accesses the local file system for models and configurations.
Sanitization: The documentation does not specify any sanitization, filtering, or validation steps for the creative content before it is processed by the LLM agents.

oransim-causal-marketing-twin