model-serving
Warn
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill includes Python scripts and documentation examples that use the
eval()function to process input within a calculator tool. - Evidence found in
examples/langchain-agents/main.pyandreferences/langchain-orchestration.md. - This allows for the execution of arbitrary Python expressions provided by the language model, which could lead to arbitrary code execution if the model's output is manipulated.
- [REMOTE_CODE_EXECUTION]: The skill documents the installation of the Ollama service via a remote shell script.
- Evidence in
examples/ollama-local/README.md:curl -fsSL https://ollama.com/install.sh | sh. - This fetches and executes code from the official Ollama website, which is a common but inherently risky installation pattern.
- [INDIRECT_PROMPT_INJECTION]: The provided agent examples ingest untrusted user data and possess capabilities that could be abused through indirect injection.
- Ingestion points: User query input in
examples/langchain-agents/main.pyand themessagefield in theChatRequestmodel inexamples/vllm-serving/main.py. - Boundary markers: Absent. The agent prompts do not include delimiters or instructions to ignore embedded commands within the processed text.
- Capability inventory: The skill includes tools for document search (reading from a database), a calculator tool using Python
eval(), and documentation for creating tools that perform raw SQL queries. - Sanitization: Absent. Input values are not sanitized or validated before being processed by the agent or its tools.
Audit Metadata