skill-iterative-loop
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from local configuration files (
/.claude-octopus/loop-config.conf) and experiment logs (/.claude-octopus/experiments/*.jsonl) to determine agent behavior and safety thresholds, which could be manipulated by local attackers. - Ingestion points: SKILL.md describes reading configuration from ~/.claude-octopus/loop-config.conf to override self-regulation weights and state from .claude-octopus/experiments/ to resume loops.
- Boundary markers: No delimiters or safety warnings are specified for the contents of these files to differentiate instructions from data.
- Capability inventory: The skill utilizes shell command execution via user-defined 'Metric' and 'Guard' commands, git operations (commit, revert), and file system writes for logging.
- Sanitization: No explicit validation or sanitization of the configuration file content or log entries is described.
- [DYNAMIC_EXECUTION]: The skill facilitates an iterative cycle where the agent is instructed to pick a change, implement it, and then execute verification commands, representing runtime generation and execution of code.
- [COMMAND_EXECUTION]: The skill allows for the execution of arbitrary shell commands through the 'Metric' and 'Guard' parameters in its automated verification mode.
Audit Metadata