suede-marketing-loops
Warn
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: MEDIUMPERSISTENCEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PERSISTENCE]: The skill instructs the agent to establish recurring tasks using system-level scheduling tools. Evidence:
SKILL.mdprovides specific instructions for wrapping workflow bodies in cron expressions (e.g.,0 9 * * 1) and verifying the installation usingcrontab -l. - [INDIRECT_PROMPT_INJECTION]: The skill ingests data from diverse external sources that could contain malicious instructions. Ingestion Points: News stories (HN, Reddit, X), social media threads, and customer feedback/reviews referenced throughout the loop catalog. Boundary Markers: Employs 'veto lists' for sensitive topics and human checkpoints for outbound actions. Capability Inventory: Orchestrates various marketing skills to draft emails, generate social media posts, and manage ad creative. Sanitization: Implements a two-tier action model that requires manual human approval for all Tier-2 actions such as spending, publishing, or sending messages.
- [DATA_EXFILTRATION]: The skill performs network operations to retrieve content from numerous external domains for analysis. Evidence: Loops defined in
references/loop-catalog.mdinvolve fetching competitor homepages, pricing data, news articles, and social media mentions.
Audit Metadata