brand-mention-monitoring
Pass
Audited by Gen Agent Trust Hub on Nov 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external YouTube and podcast transcripts through the Arcmira search and monitoring tools.
- Ingestion points: Data enters the agent's context via the
arcmira.searchandarcmira.momentumfunctions within the code blocks inSKILL.md. - Boundary markers: There are no explicit delimiters or warnings provided in the instructions to the agent regarding the potential for embedded instructions in the transcript quotes.
- Capability inventory: The skill possesses the ability to execute code (
arcmira_execute_read), write data/monitors (arcmira_execute_write), and send feedback (arcmira_feedback). - Sanitization: No sanitization or filtering of the external transcript content is implemented beyond a simple string slice (
c.text.slice(0, 300)) for display purposes. - [DYNAMIC_EXECUTION]: The skill provides JavaScript-like code templates that the agent is instructed to populate and execute using the
arcmira_execute_readtool. - Evidence: The
SKILL.mdfile contains a 'Worked program' section with logic using anarcmiraSDK-like object. This is a standardized pattern for skills utilizing the Arcmira platform to perform multi-step data retrieval and processing.
Audit Metadata