content-brief
Pass
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill instructions and metadata are consistent with its stated purpose of generating SEO content briefs. It exclusively uses the Semrush MCP for data retrieval and does not request or execute any high-risk system commands or file operations.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests external search data and user inputs to generate reports, which is a standard surface for indirect prompt injection. However, the skill implements strong defensive practices by requiring the agent to label the provenance of all data points (e.g., 'Semrush data' vs 'Model inference') and to maintain strict separation between evidence and editorial guidance. The tool's capabilities are restricted to information processing, minimizing the impact of potential injection attempts.
- Ingestion points: External keyword metrics and organic search rankings retrieved via Semrush MCP tools; user-provided domain and audience notes in SKILL.md.
- Boundary markers: Explicit instructions to use provenance labels (e.g., 'Semrush-supported guidance', 'Manual observation') and separate rationale notes for model judgments.
- Capability inventory: Research-focused MCP tools (
keyword_research,phrase_this, etc.); no capabilities for file modification, network exfiltration to unknown domains, or shell command execution. - Sanitization: Instructions to screen keyword rows for broken encoding, misspellings, and off-topic terms.
Audit Metadata