assessing-vector-and-embedding-weaknesses

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

• Third-party content exposure detected (high risk: 0.75). The runtime path scripts/agent.py → cmd_injection calls client.search(...) and then reads outsider-authored free text from h.payload["text"] (retrieved chunks from the vector store) into the LLM context via the agent’s processing/printing, enabling indirect prompt-injection content to be ingested.