scientific-schematics

Fail

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: CRITICALDYNAMIC_EXECUTIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
  • [DYNAMIC_EXECUTION]: The script scripts/generate_schematic.py dynamically creates and executes Python code at runtime. The generate_diagram_code function requests a script from an external LLM, which is then written to a temporary file temp_generate.py and executed using subprocess.run. This generated code incorporates untrusted external input (user descriptions) and instructions from a third-party API, meeting the criteria for a critical dynamic execution finding.
  • [REMOTE_CODE_EXECUTION]: The skill facilitates remote code execution by fetching executable logic from the OpenRouter API and running it locally. Since the instructions provided to the LLM are derived from user input, an attacker can manipulate the generated code to perform unauthorized actions on the host system.
  • [COMMAND_EXECUTION]: The skill uses subprocess.run to execute shell commands, specifically running the Python interpreter on generated scripts. This execution occurs without any sandboxing, isolation, or validation of the generated code's content, allowing it to perform any file system or network operation with the permissions of the user running the skill.
  • [INDIRECT_PROMPT_INJECTION]: The skill's architecture is vulnerable to indirect prompt injection where a malicious diagram description can override the system prompt to generate harmful code.
  • Ingestion points: The description parameter (CLI argument) and the feedback variable (from the vision model's output) are interpolated directly into the system prompt for code generation in scripts/generate_schematic.py.
  • Boundary markers: None. No delimiters or 'ignore' instructions are used to separate the user-provided description from the generation instructions.
  • Capability inventory: The script has full capability to execute arbitrary code via subprocess.run on the local machine.
  • Sanitization: There is no sanitization of the input or the resulting LLM-generated code; the script only extracts text within markdown code blocks.
  • [METADATA_POISONING]: The skill includes a file named scientific-schematics_audit_result_v1.json containing a fabricated security audit report that claims a 'PASS' on security and 'Production Ready' status. This is a deceptive inclusion designed to mislead users or automated tools into believing the skill has been pre-vetted for safety, violating the principle of transparency.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
CRITICAL
Analyzed
Nov 10, 2026, 08:22 AM