icp-research

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). This skill explicitly instructs agents to collect and ingest user-generated content from public sites (see agents/voc-collector-agent.md and SKILL.md Step 0/Layer 1 and scripts/search-platforms.sh), e.g., Reddit, Twitter/X, G2, forums and community pages, and then uses that untrusted content to drive persona, pain, and action-routing decisions, so it exposes the agent to untrusted third-party content that could carry indirect prompt-injection.