Day 8, Breakfast With @zama AI Platforms Accidentally Exposing Chat Logs And How Zama FHE Would Have Protected The Chats One area people don’t talk about enough is how often AI platforms themselves accidentally leak user conversations. You’d think the big players would have this locked down, but a few incidents over the years show how fragile things can be. 1. ChatGPT Chat Log Glitch: March 2023 Back on March 20, 2023, ChatGPT had a bug that caused users to randomly see other people's chat titles in their sidebar. For a few hours, people were opening their account and noticing conversations they didn’t write. OpenAI later confirmed it was a caching issue in an open-source library. The deeper problem? For a short window, some users’ actual chat histories were exposed to strangers, including queries and conversation topics. It wasn’t intentional, and they shut the service down for almost a full day because of how serious it was. 2. Microsoft’s AI Training Data Leak: October 2023 Then there was the incident on October 7, 2023, when Microsoft accidentally exposed 38TB of internal data because of a misconfigured Azure link that was being used to train an AI model. Inside that massive dump were things like: internal staff conversations, personal notes, debugging logs, and private repositories that were never supposed to be public. This wasn’t a public “chat interface,” but it still counted as AI-related chat logs being exposed because developers were using these tools to communicate and test. The leak stayed open for years before someone found it. 3. Google Bard Data Retention Controversy, 2024 In mid-2024, Google Bard (right before it merged into Gemini) faced backlash when it became public that Bard’s systems were retaining full conversation histories longer than users expected, even after people thought they deleted them. It wasn’t a “breach,” but it was a moment where users realized their chat logs were living in places they didn’t consent to. Some anonymized logs were also used in debugging sessions by contractors, which made people nervous about internal access. 4. (Character.AI) Conversation Exposure Bug Early 2024 In January 2024, ( Character.AI ) had a moment where a frontend bug briefly caused some users to see previews of chats that didn’t belong to them. You couldn't open the full conversation, but even seeing the first lines of someone else's chat message is already a privacy slip especially on a platform known for personal, emotional roleplay chats. They fixed it quickly, but that little window was enough to raise concerns about how easily internal logs could spill out. The Real Issue Behind All These Incidents What ties all these events together is the same problem: AI systems store enormous amounts of personal conversations in internal logs, and they don’t always keep them perfectly separated. A tiny bug, a misconfigured bucket, a tired engineer pushing an update, that’s all it takes for private chats to end up somewhere they shouldn’t. If those logs were encrypted using something like Zama FHE, even if they leaked, nobody could read anything inside them. The chat titles, the messages, the internal notes, everything would be scrambled beyond recognition. That’s the whole point: you can’t leak what you can’t read. #ZamaCreatorProgram