delta-science

Pass

Audited by Gen Agent Trust Hub on Nov 11, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill includes Python scripts (scripts/invoke.py, scripts/catalog.py) that facilitate the execution of the delta-cli binary using subprocess.run. While the execution is structured (passing arguments as a list) and intended for tool orchestration, it represents a core capability for interacting with the host system.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external sources such as PubChem synonyms, PubMed literature summaries, and protein annotations. This data is used to generate reports and can influence subsequent tool calls, which could be exploited if an attacker poisons these external databases with malicious instructions designed to manipulate the agent's behavior.\n
  • Ingestion points: Data is ingested via the delta-cli science invoke output, which pulls from remote scientific services and databases (e.g., PubChem, PubMed, Huada Protein Tools).\n
  • Boundary markers: The skill explicitly instructs the agent to use evidence boundary text ("Evidence boundary: business results in the current Delta CLI data only") and warns against adding information absent from the response.\n
  • Capability inventory: The skill has the capability to write to the workspace, execute the delta-cli binary with various parameters, and perform multi-step scientific workflows.\n
  • Sanitization: The skill relies on natural language instructions for the agent to strictly follow schemas and ignore non-scientific content, but does not implement programmatic sanitization of the data retrieved from external APIs before it enters the prompt context.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 11, 2026, 02:06 AM