complexa-sweep

Warn

Audited by Gen Agent Trust Hub on Aug 13, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute shell commands using variables directly gathered from user dialogue, creating a potential command injection surface.\n
  • Evidence: Step 4 in SKILL.md shows the construction of shell commands and loops (e.g., python script_utils/generate_inference_configs.py and complexa design) using the $run_name and $pipeline_config_name variables.\n
  • Risk: Maliciously crafted user input containing shell metacharacters could lead to arbitrary command execution as the inputs are not explicitly sanitized in the instructions.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data from external CSV files and user-defined parameters without adequate boundary markers.\n
  • Ingestion points: User-provided parameters (run_name, task_name) and external results_*.csv files parsed in Step 5.\n
  • Boundary markers: Absent; no instructions are provided to the agent to treat this data as untrusted or to ignore embedded instructions.\n
  • Capability inventory: The skill has Bash, Read, and Write capabilities, which could be exploited if the ingested data contains malicious instructions.\n
  • Sanitization: Absent; the skill does not specify any filtering or validation of the CSV content before using it to generate summary reports and manifests.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 13, 2026, 07:51 AM
Security Audit — agent-trust-hub — complexa-sweep