auto-review-loop-minimax

Pass

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: SAFECREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill instructions specify reading from ~/.codex/settings.json to retrieve the MINIMAX_API_KEY. Accessing configuration files containing secrets outside the immediate project environment is a sensitive operation.
  • [INDIRECT_PROMPT_INJECTION]: The skill implements an autonomous loop that processes external data as instructions.
  • Ingestion points: The skill ingests research feedback from the MiniMax API (https://api.minimax.io/v1/chat/completions).
  • Boundary markers: No delimiters or safety instructions are provided to the agent to help it distinguish between legitimate feedback and potentially malicious instructions embedded in the API response.
  • Capability inventory: The skill has access to file writing tools, shell execution (via curl and bash), and remote server access via ssh to GPU nodes.
  • Sanitization: There is no mechanism described for validating or sanitizing the content returned by the external API before the agent implements changes or executes code based on it.
  • [COMMAND_EXECUTION]: The skill uses curl to interact with an external API and ssh with screen/tmux to manage long-running processes on remote servers. It also includes a rule for the agent to execute cat commands via bash without seeking user permission to handle large file writes.
  • [DYNAMIC_EXECUTION]: The agent is instructed to write and modify experiment scripts and then execute them, which involves generating and running code at runtime based on the autonomous loop's state.
  • [EXTERNAL_DOWNLOADS]: Fetches content from the MiniMax API. While this is a well-known service used for its intended purpose, the automated integration of its responses into a development workflow represents an external data dependency.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 10, 2026, 09:36 AM