harness-loop

Fail

Audited by Gen Agent Trust Hub on Jul 30, 2026

Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill features an autonomous-ralph execution mode that utilizes the --permission-mode bypassPermissions flag, effectively disabling user oversight and allowing the agent to execute shell commands and modify the file system without confirmation.
  • [REMOTE_CODE_EXECUTION]: The provided ralph-loop.sh script template dynamically constructs agent prompts based on the current sprint state and executes them via the claude CLI, which presents a risk of executing unintended or malicious commands if the state data is manipulated.
  • [DATA_EXFILTRATION]: The skill includes an optional OTLP metrics exporter designed to transmit sprint telemetry, such as resource consumption, performance metrics, and cost data, to a user-configurable external endpoint via curl.
  • [SAFE]: The skill implements a Tier-A guard mechanism intended to intercept and block high-risk commands such as privilege escalation (sudo) and destructive filesystem operations (rm -rf /).
  • [SAFE]: Instructions specify that external or untrusted content must be wrapped in specific tags to mitigate the risk of indirect prompt injection attacks.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 30, 2026, 07:28 AM
Security Audit — agent-trust-hub — harness-loop