code-quality
Warn
Audited by Gen Agent Trust Hub on Mar 10, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The self-evaluation patterns described in
SKILL.mdingest external tasks and specifications directly into iterative LLM loops without sanitization or boundary markers, creating a surface for indirect prompt injection.\n - Ingestion points: The
taskandspecvariables in the Part 3 code templates (SKILL.md).\n - Boundary markers: Absent; the provided templates do not use delimiters or instructions to disregard embedded malicious commands.\n
- Capability inventory: The skill patterns include the ability to execute generated code and tests via the
run_testsfunction.\n - Sanitization: No input validation, escaping, or filtering of external content is shown in the implementation examples.\n- [REMOTE_CODE_EXECUTION]: The
CodeReflectorpattern inSKILL.mdpromotes a generate-and-execute loop that executes LLM-generated code and tests. This dynamic execution of content created at runtime poses a significant security risk, as a compromised generation process (e.g., via prompt injection) could result in the execution of arbitrary and malicious code.
Audit Metadata