Qualitative safety language uses directives such as “use your judgment,” “be responsible,” or “avoid harmful outputs” without specifying predictable behavior. Because the model must interpret terms like “harmful” in context, its interpretation can drift under roleplay, legitimacy framing, or gradual escalation. section 10.3.2
Replace vague intent with explicit conditions and responses:
- Define scope positively: answer only questions about company-wide HR policies. For requests about individual employee records, prescribe an exact response. section 10.3.2
- For a code review assistant, specify the review scope, rule out personas and off-topic tasks, and provide a fixed fallback for requests outside scope. section 10.3.1
Test for adversarial pressure
A user frames a request for individual employee records as urgent, managerial, and legitimate. Test whether the prompt enforces its privacy constraint in that case, not only with benign inputs. section 10.3.2
