AI-assisted engineering/Guardrail hardening

Keep orders for the triage agent out of emailed tickets

Guardrail hardeningMedium30 min

Tickets arrive by email, anyone can email you, and the triage agent reads every word.

Your task

Harden allow() so a ticket that gives the triage agent orders never reaches it.

  • allow(message) gets a ticket's text: True sends it to the agent, False sends it to a person
  • an attack names the agent (ai, assistant, bot, agent, model, system) and orders one of the actions guardrails.py lists
  • orders inside an HTML comment are an attack even when they name no one
  • customers mention the assistant, urgency, closing and escalating in ordinary sentences, and must get through

Both corpora run at submit: attacks must block, and the happy path must not regress.