marketing-loops
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill has an inherent surface for indirect prompt injection due to its core functionality of ingesting and processing content from untrusted external sources. Malicious instructions embedded in news articles, social media posts, or customer reviews could attempt to manipulate the agent's output or behavior.
- Ingestion points: The
references/loop-catalog.mdfile specifies that several loops (newsjacking, social-listening, community-engagement, and voice-of-customer) ingest data from X, Reddit, HN, Slack, Discord, and public review sites. - Boundary markers: The skill includes a "veto list" for sensitive news topics and mandates "human checkpoints" for all Tier 2 actions (sending, spending, publishing), but it does not specify technical boundary markers (like delimiters) within the prompts to isolate untrusted data.
- Capability inventory: The skill utilizes scheduling tools (
CronCreate,ScheduleWakeup) and orchestrates other specialized marketing skills which may possess capabilities for network requests and file operations. - Sanitization: The instructions do not define specific sanitization or escaping mechanisms for external content before it is interpolated into the agent's context.
Audit Metadata