ui-designer — agentic threat model

Utilizes vision-language models to analyze UI screenshots. Vulnerable to adversarial image perturbations (visual prompt injection) that could manipulate the extracted design system or inject malicious instructions into the generated output prompts.

Not certain from the listing — processes image inputs and writes design-spec outputs. There is no explicit mention of vector databases, RAG, or persistent storage, suggesting a transient data lifecycle with minimal data operations risk.

Orchestration is limited to a simple pipeline: reading an image, extracting features, and writing a text specification. Low risk of tool misuse or framework vulnerabilities due to the absence of active execution tools.

Not certain from the listing — as an open-source community skill, deployment depends entirely on the host environment. Standard containerization and input size limits are recommended to prevent denial-of-service via resource-intensive image processing.

Not certain from the listing — no built-in evaluation, logging, or guardrails are described. Lacks mechanisms to detect if the generated design prompts contain injected malicious payloads or system instructions.

Not certain from the listing — no identity, access control, or compliance frameworks are specified. Relies on the parent platform to enforce data privacy, especially if users upload proprietary or sensitive mockups.

Operates as an isolated, single-purpose skill. While listed in a community directory, it does not natively initiate multi-agent coordination or autonomous marketplace transactions, minimizing ecosystem risks.