LLM02:2025 — Sensitive Information Disclosure

Slide 17 · The Pattern

Three different mechanisms. One shared root cause.

🔓

Unintentional Exposure (OpenAI Redis bug)

Infrastructure didn't verify ownership before returning cached data.

🎯

Targeted Injection (Slack AI)

The model's retrieval and rendering behavior was redirected by planted text.

📚

Training Data Leak (poem attack)

Sensitive content was never removed before it became part of the model's weights.

The Common Thread

In every case, sensitive data and ordinary data were treated identically — by a cache, by a retrieval system, by a training pipeline. None of these systems asked “should this specific piece of data be visible to this specific requester?” That question, asked everywhere data moves, is what Part 4's mitigations are really about.

← Back Next → Prevention Overview