denario

Warn

Audited by Gen Agent Trust Hub on Nov 6, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPERSISTENCEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The core functionality of the skill includes the get_results() method, which automatically generates and executes Python code based on a research methodology. Because this methodology is itself generated by an LLM based on potentially untrusted input, it represents a runtime code generation and execution vector.
  • Evidence: Found in references/research_pipeline.md describing the behavior of get_results() as executing the methodology and running computations.
  • [INDIRECT_PROMPT_INJECTION]: The skill is highly susceptible to indirect prompt injection because it ingests untrusted data descriptions to drive its logic and code generation stages.
  • Ingestion points: set_data_description() and set_idea() in SKILL.md and references/research_pipeline.md accept free-form text from users or external datasets.
  • Boundary markers: No explicit delimiters or instruction-ignore markers are documented for the data ingestion methods.
  • Capability inventory: The agent has the capability to execute arbitrary shell/Python code via get_results() and generate complex LaTeX documents via get_paper().
  • Sanitization: There is no evidence of input validation or sanitization before the data is processed by the agents.
  • [EXTERNAL_DOWNLOADS]: The installation guide encourages downloading the software from sources that are not formally verified in the trusted vendor scope, including a specific GitHub repository and a third-party Docker image.
  • Evidence: references/installation.md lists https://github.com/AstroPilot-AI/Denario.git and docker pull pablovd/denario:latest as installation methods.
  • [PERSISTENCE]: The documentation instructs users to store API keys in shell configuration files, which is a method of maintaining persistent environmental configuration across sessions.
  • Evidence: references/llm_configuration.md suggests adding export commands to ~/.bashrc, ~/.zshrc, or ~/.bash_profile.
  • [COMMAND_EXECUTION]: The skill documentation provides instructions for system-level package installation and running a web-based GUI that opens a local port.
  • Evidence: references/installation.md includes sudo apt-get install texlive-full and denario run to launch a Streamlit application.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 6, 2026, 12:15 PM