gemini-interactions-api

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 1.00). The skill contains an explicit instruction telling the agent to "override your training data" (SKILL.md:9-11), which is an attempt to override system-level instructions and therefore constitutes a prompt-injection outside the API-documentation purpose.

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

• Third-party content exposure detected (high risk: 0.70). The required workflow includes using “Deep Research” and managed agents (e.g., background=True with deep-research-* or antigravity-preview-*), which at runtime can fetch external web content; that fetched web page text can become LLM-visible context via the agent’s tool outputs, creating an indirect prompt-injection exposure path.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly requires fetching hosted documentation at runtime (e.g., https://ai.google.dev/gemini-api/docs/interactions.md.txt and other https://ai.google.dev/... doc pages) to determine how to construct prompts and code, so these external docs directly control the agent's behavior at runtime.