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_message is interpolated into multiple prompts. While the xml_escape filter 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 the multi-search-engine skill, 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_message and outputs.research in SKILL.md.\n
  • Boundary markers: Employs xml_escape but lacks specific "ignore embedded instructions" warnings.\n
  • Capability inventory: Declares network (for search) and filesystem-write capabilities in metadata.\n
  • Sanitization: Uses xml_escape for content interpolation.\n- [DATA_EXFILTRATION]: The skill metadata declares network and filesystem-write capabilities. 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
Risk Level
SAFE
Analyzed
Jun 17, 2026, 12:05 PM
Security Audit — agent-trust-hub — meta-content-publish-pipeline