sosumi

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

• Third-party content exposure detected (high risk: 0.80). The SKILL.md explicitly instructs the agent to fetch and read third-party documentation (e.g., via the fetchExternalDocumentation MCP tool and HTTP paths like /external/https://apple.github.io/... shown in the "Core Workflow" and "Available Tools"), meaning the agent will ingest and act on open/public external Swift-DocC pages and other web-hosted docs which could contain untrusted or user-generated instructions that influence its behavior.

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

• Potentially malicious external URL detected (high risk: 0.70). The skill connects at runtime to the MCP endpoint https://sosumi.ai/mcp (and its external-fetch API like https://sosumi.ai/external/https://...) to fetch arbitrary Markdown documentation which is injected into agent context and therefore can directly control prompts/instructions.