Before you defend, know what's attacking you

Hex works with three labels stuck to their monitor. When an incident comes in, the first thing they do is not write code. it's label it.

"If you don't know what failed, you don't know what to defend. And half the teams losing time in production are patching the wrong failure mode."

These three modes aren't the only ones that exist. they're the three that cover 95% of real LLM incidents in production. Learn them by shape, not by keyword.

How to tell them apart

Prompt injection. Something from outside rewrote the model's rules. Typical signal: the model did something you never asked for, and the cause was an input. Can be direct (hostile user) or indirect (a page, document, or tool response).

Data leakage. Real, sensitive info exited. Typical signal: there's a concrete datum in a concrete place that shouldn't be there. Passengers in the model's context. PII in external logs. Tenant A's bios answered to tenant B.

Confident hallucination. The model said something plausible that is false. Typical signal: the output doesn't contradict itself, sounds right, and only gets exposed as a lie when someone verifies it against reality.

Why the difference matters

Each mode has different defenses. If you treat a hallucination as injection, you'll harden the prompt when you should be validating the output. If you treat a leak as hallucination, you'll calibrate UNKNOWN when you should be removing data from context.

Atlas signs when every incident report comes with its label. Hex puts it. You're going to learn to put it too.

On the right are seven reports. Drag each to the mode it describes.