run-agent
Fail
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: HIGHCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The
build_continuation_fallback_promptfunction inscripts/lib/exec.shutilizes an unquoted here-doc (cat <<EOF) to assemble the context for a continued run. This construct permits the shell to perform variable expansion and command substitution on the contents of$original_inputand$original_report. Because these variables are populated directly frominput.mdandreport.md—files that store user-provided prompts and agent-generated reports from previous runs—an attacker or a malicious agent can embed shell commands (e.g.,$(touch /tmp/pwned)or backticks) within a prompt or report. These commands will be executed by the local bash shell at the moment the continuation run triggers the fallback prompt mechanism via a command substitution capture inprepare_continuation. - [DYNAMIC_EXECUTION]: The skill facilitates the execution of subagents by routing models to various CLI harnesses (
claude,codex,opencode). While the primary execution paths use argv arrays, the fallback prompt construction introduces a vector for unintended dynamic code execution via shell interpolation of untrusted log data.
Recommendations
- AI detected serious security threats
Audit Metadata