AI Security Issues Are Social Engineering Issues

Modern transformers have introduced prompt injection as a new class of security weakness in LLMs. LLMs follow instructions in content, allowing an instruction to be snuck into inputs. Capability-Based Security is part of the fix.