Slide 10 of 27
Part 2 · TypesSlide 10
Slide 10 · PII Leakage — Real Example
The OpenAI Redis Bug — March 20, 2023.
Confirmed Incident · March 20, 2023 · OpenAI's Own Disclosure
ChatGPT Bug Exposes Other Users' Chat Titles and Payment Data
No CVE · Root Cause: redis-py library bug · Disclosed by OpenAI, March 24, 2023

The setup: ChatGPT used the open-source redis-py library to cache session data. A bug in that library meant cancelled requests could corrupt connections, returning unrelated data from the cache.

What leaked: During a nine-hour window, some users could see chat titles from another active user's history. Separately, OpenAI disclosed that 1.2% of ChatGPT Plus subscribers active during that window may have had billing information exposed — first and last name, email, payment address, the last four digits of a card, and expiration date.

The fix: OpenAI patched the library bug and added redundant checks to confirm cached data actually belongs to the requesting user.

Why it matters for LLM02: this is OWASP's own “Unintentional Data Exposure” scenario, observed in production — no attacker required. A caching bug in unrelated infrastructure was enough to cross the privacy boundary between two users.
The Defense This Would Have Stopped

Robust validation that returned data actually belongs to the requesting session — OWASP's “Robust Input Validation” mitigation, applied to caching layers, not just user prompts.

← BackNext → Proprietary Algorithm Exposure