AI guardrails: strategies and options

What was covered

• The core problem. AI projects rest on probability distributions, not fixed rules, so the same input does not always give the same output. That leaves a reliability gap to manage, especially where reproducibility and critical logic matter.
• Deterministic against probabilistic. A rule-based system is fully predictable. A probabilistic one picks the likeliest outcome. A guardrail is a deterministic control layer that imposes a fixed branch on the probabilistic solution.
• Three reasons to have one. Security, cost control (filtering expensive LLM calls) and compliance, including GDPR and the EU AI Act.
• A live demo on Gemini 2.5 Flash, five scenarios. A normal request passes. A prompt injection is blocked before any tokens are spent. Personal data is redacted before it reaches the model. Two off-topic requests (how to fry an egg, a cover letter for another employer) are blocked to save tokens.

What came up

• Test differently. Because outputs vary, build an evaluation harness instead of relying on unit tests.
• A new attack surface. Prompt injection needs particular care in architectures where an LLM can execute code, because a single success could compromise a whole company.
• Redact before you store. Companies are adopting a personal-data reduction layer before saving data, to meet regulations such as GDPR.
• Practice. Gandalf AI is an online game for trying prompt injection yourself.

Takeaway

For enterprise applications, guardrails are not optional. They are a basic part of the system.