Feels like a variant of the “prompt injection” vulnerabilities that are so amusing in LLMs. Call it “knowledge graph injection”. Perhaps it’s a fitting penalty

Feels like a variant of the “prompt injection” vulnerabilities that are so amusing in LLMs. Call it “knowledge graph injection”. Perhaps it’s a fitting penalty for the seemingly common practice of making support numbers difficult to find (or worse, nonexistent)