experiment-queue
Warn
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The scheduler script
scripts/queue_manager.pyconstructs shell commands for execution on remote hosts usingsubprocess.run(shell=True)without consistent sanitization of all inputs. - Evidence: The
runfunction inscripts/queue_manager.pyis a thin wrapper aroundsubprocess.run(cmd, shell=True), which is used for multiple sensitive operations. - Evidence: In
launch_job, several variables derived from the user-provided manifest, such asjob['id']andconda_env, are interpolated into command strings without sufficient sanitization. For example,screen_name = f"EQ_{job['id']}"is used directly inscreen -dmS {screen_name}without quoting, allowing for command injection ifjob['id']contains shell metacharacters. - [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves processing external data (grid specifications and manifests) to automate experiment execution, which creates an entry point for malicious instructions.
- Ingestion points:
scripts/build_manifest.pyreads user-provided YAML/JSON files, andSKILL.mdinstructions guide the agent to parse natural language descriptions into a manifest. - Boundary markers: Absent. There are no explicit delimiters or instructions to ignore embedded commands within the processed configuration data.
- Capability inventory: The skill possesses the capability to execute arbitrary shell commands on remote servers, write files, and perform network operations via
sshandscp. - Sanitization: Partial. While
shlex.quoteis used for some path-related variables, other fields in the manifest are used in shell commands without proper escaping. - [DYNAMIC_EXECUTION]: The skill dynamically generates and executes complex shell command strings at runtime, increasing the risk of injection vulnerabilities.
- Evidence: The
launch_jobfunction inscripts/queue_manager.pyassembles a multi-line bash script (including directory changes, environment activation, and environment variable setting) which is then passed toscreenandbash -cfor execution. - [EXTERNAL_DOWNLOADS]: The skill automates the transfer and execution of code on remote servers as part of its core functionality.
- Evidence: Step 3c in
SKILL.mdusesscpto copy helper scriptsqueue_manager.pyandbuild_manifest.pyto the target server, which are subsequently executed viasshin Step 3d.
Audit Metadata