ai-rag

Fail

Audited by Gen Agent Trust Hub on Oct 17, 2026

Risk Level: CRITICALINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of data from external sources such as web crawls, documents, and tool outputs to augment LLM responses. This process exposes the system to indirect prompt injection, where malicious instructions within the retrieved content could attempt to override the agent's behavior. \n
  • Ingestion points: Data enters the context through document loaders and ingestion pipelines described in SKILL.md and references/research-and-ingestion-patterns.md. \n
  • Boundary markers: The skill employs clear delimiters in its context packaging templates (assets/context/template-context-packing.md) to distinguish between instructions and retrieved evidence. \n
  • Capability inventory: The skill utilizes network requests for source validation (scripts/check_sources.py) and local command execution for testing (scripts/test_eval_regressions.py). \n
  • Sanitization: Defensive strategies are documented in references/security-red-team-cases.md, including escaping HTML and using LLM-based reasoning judges to verify faithfulness. \n- [COMMAND_EXECUTION]: The skill includes a testing utility, scripts/test_eval_regressions.py, which uses the subprocess module to execute other internal Python scripts within the skill's repository. These operations are intended for internal validation and regression testing of the retrieval and evaluation logic. \n- [EXTERNAL_DOWNLOADS]: The skill references and provides URLs for various research papers and development tools, including the TruLens observability framework. While an automated scan flagged the TruLens domain, it is an established tool within the AI research community used here for legitimate evaluation and instrumentation purposes.
Recommendations
  • Contains 1 malicious URL(s) - DO NOT USE
Audit Metadata
Risk Level
CRITICAL
Analyzed
Oct 17, 2026, 08:31 AM