superforecast

Pass

Audited by Gen Agent Trust Hub on Nov 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill demonstrates a potential surface for indirect prompt injection by ingesting untrusted user data (forecast questions, evidence, and reasoning) and using the Bash tool to pass this data as arguments to a local Python script.
  • Ingestion points: User-provided questions, canonical phrasings, evidence descriptions, and reasoning strings defined in SKILL.md workflows and the eight gates of the forecasting process.
  • Boundary markers: The instructions do not specify the use of boundary markers, such as XML tags or special delimiters, to separate user-provided content from instructions during tool invocation.
  • Capability inventory: The skill extensively uses the Bash tool to run the scripts/sf.py engine. It also has access to Write and Edit tools to modify files in the global ledger located at ~/.superforecast/.
  • Sanitization: There is no evidence of sanitization or escaping logic in the instructions to ensure that shell metacharacters (e.g., backticks, dollar signs) in user input do not trigger arbitrary command execution when the agent invokes the script via the shell.
  • [COMMAND_EXECUTION]: The skill's operational model relies on the agent executing shell commands to interact with a deterministic Python engine. While the provided script scripts/sf.py is safe and uses only standard Python libraries, the pattern of passing arbitrary user-provided strings into shell commands (e.g., python3 sf.py new "<user_input>") creates an inherent risk of command execution if the input is not carefully handled by the agentic runtime.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 21, 2026, 06:47 PM