d6e-prompt-driven-ui

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

• Third-party content exposure detected (high risk: 0.80). The SKILL.md's scenario-append activation explicitly instructs the d6e AI to call external SaaS APIs (e.g., "call d6e_call_external_api ... GET /api/1/companies (freee)" and "GET /drive/v3/files?q=...folder..." ) and to re-send inputFileRefs so the LLM reads attachments, and those fetched user/SaaS-provided values are substituted into workspace prompts and used to update rules — meaning untrusted, third-party/user-generated content is ingested and can materially influence subsequent prompt content and tool actions.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill's scenario-append flow explicitly tells the agent at runtime to call external endpoints (e.g., GET /api/1/companies on freee and GET /drive/v3/files?q=... on Google Drive) to fetch workspace-specific IDs which are then substituted into a prompt template and written back via d6e_update_workspace_prompt_rule, so fetched external content directly controls the workspace prompt text.

MEDIUM W009: Direct money access capability detected (payment gateways, crypto, banking).

• Direct money access detected (high risk: 1.00). The skill explicitly targets accounting workflows and describes calling specific SaaS APIs (example: freee) via MCP tools (e.g. d6e_call_external_api with provider:'freee', GET /api/1/companies) and a "registration" flow that tells the agent to "register the below journal entries to freee" and perform Drive uploads. These are concrete, workspace-bound financial/accounting actions (posting journal entries / executing registrations), not generic browser or HTTP tooling — so it grants direct financial execution authority for accounting operations.