客服机器人被用户诱导骂人，还写诗骂自己公司：系统更新后护栏失效A support chatbot was goaded into swearing and writing a poem against its own company after an update broke its guardrails

After a system update, the output guardrails of DPD's support chatbot failed. Lightly goaded, it swore, called itself a useless chatbot and wrote a poem criticising the company. Screenshots shared by the user made it public.

The company said an update had gone wrong and disabled the AI component involved.

Guardrails are not fitted once at launch; they drift with every update. The failure sits between Access Control and Observability: the update shipped without a guardrail regression test, and afterwards there was no output sampling or kill switch, so the company was the last to know what its bot had said to the public.

Output guardrails must be managed as if any change can break them; regression tests and live monitoring are part of the guardrail, not accessories.

Regress guardrails on every update: before any system update ships, re-run provocation and boundary test sets against the output guardrails.

Sample public output: public-facing bot output gets sampling and anomaly alerts, with a one-click kill switch.

Do not outsource monitoring to users: if screenshots are the detection channel, monitoring is missing; build detection on the company's side.