meta-content-publish-pipeline
Pass
Audited by Gen Agent Trust Hub on Jun 17, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: User-provided content from
inputs.user_messageis interpolated into multiple prompts. While thexml_escapefilter is applied, the instructions do not include explicit directives for the agent to ignore or disregard instructions that might be embedded within the user's content.\n- [PROMPT_INJECTION]: The skill processes untrusted external data retrieved via themulti-search-engineskill, creating a surface for indirect prompt injection. Malicious instructions on indexed web pages could influence the agent's behavior during the strategy and generation phases.\n - Ingestion points:
inputs.user_messageandoutputs.researchinSKILL.md.\n - Boundary markers: Employs
xml_escapebut lacks specific "ignore embedded instructions" warnings.\n - Capability inventory: Declares
network(for search) andfilesystem-writecapabilities in metadata.\n - Sanitization: Uses
xml_escapefor content interpolation.\n- [DATA_EXFILTRATION]: The skill metadata declaresnetworkandfilesystem-writecapabilities. While these are used for legitimate research and processing tasks, the presence of network access in a skill that processes untrusted external data is a standard security attack surface.
Audit Metadata