The Agent Skills Directory

Instructional Override Directives: The skill contains instructions that explicitly direct the agent to override its training data in favor of the provided content. This pattern is often used in technical skills to ensure the use of the most current API versions, but it employs techniques similar to prompt injection to modify agent behavior.
Indirect Prompt Injection Surface: The skill facilitates the creation of agents that ingest external data and possess broad capabilities such as shell access and web browsing, which creates a potential surface for indirect prompt injection.
Ingestion Points: Data enters the system through the input parameter in interactions.create (SKILL.md) and repository source URLs in agents.create (SKILL.md).
Boundary Markers: The provided implementation examples do not include explicit delimiters or instructions for the agent to ignore potentially malicious commands within processed data.
Capability Inventory: The documented antigravity agent supports code execution (Bash, Python, Node.js), file management, and web access.
Sanitization: The skill does not demonstrate specific sanitization or input validation for data processed by these agents.
Remote Dynamic Execution: The API documented allows for the dynamic execution of code within managed, sandboxed environments. While these are hosted services provided by the vendor, the pattern of executing runtime-generated code based on external input is a security consideration for developers using these tools.

gemini-api-dev