An agent reads text written by strangers and can call tools that change things. An instruction in a prompt is a request the model usually follows. A guardrail is a condition that code enforces whether the model follows it or not. This site starts from the validations, permissions and business rules of ordinary backend systems and climbs, one idea at a time, to layered controls around a probabilistic agent, asking one question at each step: what stops the harm when the model is wrong?
A roof-assessment description carries a hidden request to export applicant records. A sentence in the prompt that says “never export data” is a hope, not a control. What actually stops it is code the model cannot argue with: a tool that is not allowed, a parameter outside its bounds, a policy check, an approval, and an audit record that shows the attempt.
Validations, access control and policy engines before AI; guardrails against instructions, evaluations and governance; the trust boundaries an agent crosses.
Type, format, personal data and tenancy at the door; allow-lists, parameter bounds and approvals on tool calls; schema, grounding and leakage checks on output.
Direct and indirect injection, sandboxing, network egress, least privilege, policy as code, and model-based guardrails with their error rates.
Escalation and override, adversarial suites and red teaming, fail-open against fail-closed, auditability and exception handling.