editor-test-harvester

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

• Third-party content exposure detected (high risk: 0.95). The skill’s runtime workflow clones/reads the target editor repository (e.g., git clone https://github.com/... ../lexical and then rg --files "$target" / rg -n ... <target-test-files> / “Read all portable and uncertain files”), so outsider-authored test source text from that repo is ingested into the agent’s LLM context for extraction/classification.

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

• Potentially malicious external URL detected (high risk: 1.00). The Discovery Workflow contains a runtime git clone command that pulls remote code—git clone https://github.com/facebook/lexical.git ../lexical—which the skill then reads to drive harvesting decisions, so the fetched content can directly control agent behavior.