codex-claude-cursor-loop

Fail

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: HIGHCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • COMMAND_EXECUTION (HIGH): The skill uses a pattern of interpolating LLM-generated content (Claude's plans and Cursor's implementation summaries) directly into shell command strings.
  • Evidence: Phase 2, Phase 5, and Phase 7 use echo "[content]" | codex exec. If the LLM content contains shell metacharacters like backticks, semicolons, or dollar signs, it can lead to arbitrary command execution on the host runner.
  • Evidence: Phase 4 and Phase 7 pass LLM content as arguments to cursor-agent. If the content contains characters that break the argument wrapping, it can lead to command injection.
  • EXTERNAL_DOWNLOADS (MEDIUM): The skill requires the external CLI tools codex and cursor-agent to be installed and available in the PATH. These are not standard system utilities and their security profile depends on the environment setup.
  • INDIRECT_PROMPT_INJECTION (LOW): The skill implements a multi-step loop where it reads and processes data generated by different agents, creating an attack surface for indirect injection.
  • Ingestion points: Phase 6: Claude's Final Review reads implemented code from the filesystem; Phase 5 reads implementation summaries.
  • Boundary markers: Absent. Content is interpolated directly into prompts and shell commands.
  • Capability inventory: cursor-agent (file system write access), codex exec (logic execution).
  • Sanitization: None. There is no escaping or validation of the strings before they are used to influence the next step in the loop.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 17, 2026, 06:07 PM