carbon

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

• Third-party content exposure detected (high risk: 0.85). SKILL.md instructs refreshing the live catalog via a runtime curl of an external URL (https://api.agent-ds.oday-bakkour.com/.../design.md), which would inject outsider-authored free-form text (the fetched DESIGN.md content) into the agent’s LLM context.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly instructs runtime refresh via curl to https://api.agent-ds.oday-bakkour.com/v1/systems/carbon/design.md, which returns a DESIGN.md containing the "Agent Prompt Guide" and token rules that directly control agent prompts and behavior.