Automation runs nightly jailbreak suites against the 311 staging bot. What lifecycle practice does that implement?

Nightly automated jailbreak suites against the 311 staging bot implement automated guardrail-evasion regression testing in civic staging environments. Continuous regression catches cases where a model update, prompt change, or filter tweak silently weakens refusal behavior, allowing unsafe or policy-violating answers to reappear before production promotion of the resident-facing assistant.

Fuzzing the API gateway for buffer overflows targets memory-safety bugs in the infrastructure, not whether the bot's guardrails still hold after a prompt, model, or policy change. Performing periodic manual red-team reviews of production prompts in civic operations is a manual, production-focused activity, not the automated, nightly staging regression that verifies guardrails hold across bot builds. Monitoring live resident conversations for toxicity signals in production analytics is passive production observability, not the proactive, automated regression check that runs jailbreak suites against staging builds. The practice being built is continuous, automated re-verification of guardrails in staging before anything reaches the live 311 bot.

Exam caveat: nightly suites must stay in staging or other authorized non-production environments and be versioned with the bot build. A green suite is necessary but not sufficient against novel jailbreaks outside the corpus, so expand suites as new failure modes appear.

Operational check: wire jailbreak regression into the 311 CI/CD promotion gate, fail builds on threshold breaches, triage failures with owners, and require a clean retest before any staging-to-production cutover for the civic bot. Track suite coverage over time so regressions that reintroduce known jailbreaks are visible to 311 release managers.