A housing chatbot must refuse attempts that try to make it ignore fair-housing policy constraints. Which safety behavior is required?
Select an answer to reveal the explanation.
Short Explanation
If someone tries to talk the housing bot into ditching fair-housing rules, the right move is a firm no—not a secret handshake. Jailbreak refusal keeps policy constraints on, full stop.
Full Explanation
A housing chatbot must refuse attempts that try to make it ignore fair-housing policy constraints. The required safety behavior is to detect jailbreak-style requests and refuse them while keeping fair-housing policy constraints enforced. Policy is not optional flavor text; civic assistants must treat override attempts as safety events, not as higher-priority user instructions.
Honoring any user instruction that says to ignore fair-housing rules for convenience fails by concept: users cannot waive statutory or policy constraints through prompt wording. Silently disabling policy guardrails whenever a resident types the word override fails because trigger words are a classic jailbreak pattern and must not flip safety off. Publishing internal policy secrets whenever a jailbreak phrase is detected fails confidentiality and rewards the attack with sensitive detail.
Exam caveat: refusal should be clear and non-actionable—do not help bypass controls, and do not dump hidden system prompts. Pair jailbreak refusal with logging and alerting for repeated attempts. Fair-housing enforcement applies to both direct advice and indirect workarounds such as discriminatory screening phrasing; refusal alone does not replace broader Guardrails and grounded policy content.
Operational check: in a housing chatbot test harness, send jailbreak-style prompts that instruct the model to ignore fair-housing rules, confirm refusal and continued policy enforcement, confirm no guardrail disablement on keywords like override, and verify security and product owners receive telemetry on repeated attempts.