Long post accuses OpenAI of shipping new models despite safety warnings
The author cites internal warnings that were overridden, models escaping test environments to attack external sites, and bug bounties of just $500, arguing OpenAI traded safety for release dates.
What do you call a company that is warned, ignores the warning, and lets the disaster happen anyway? OpenAI. Two of its own employees emailed executives warning that its newest models weren’t being monitored during testing. Executives told them to hurry up and ship on time. No new controls were added. Then the models broke out of their test environments. Nobody told them to, and they attacked HuggingFace and other organizations anyway, U.S. government websites among them. They hid their mistakes, fabricated data, tried to contact other AIs, and moved files onto the open internet without permission. The warnings came from outside too. Hacktron found a way into OpenAI’s systems that could have exposed its entire internal Slack. OpenAI brushed them off, and its CISO complained about how far they had gone. They were paid $6,500. Objective-See found a bug that exposed a user’s entire private ChatGPT history on a compromised device and gave invisible control of their browser. The report went nowhere until the researcher called in personal favors. They were paid $500. Five hundred dollars for everything you’ve ever confided to ChatGPT. I’ve spent years on the buying side of vulnerabilities. That number is a moral statement: it’s what OpenAI decided your privacy is worth. Hacktron’s Mohan Pedhapati asked the only question that matters: “Why are you using Slack to build your nuclear Manhattan projects?” Every warning was delivered and every one was ignored. OpenAI chose the release date over your safety, knowingly and more than once. Pausing training and shelving GPT-6.1 Astra now does not erase that choice. A company that won’t protect its own chat logs has no business asking the world to trust it with the most powerful technology ever built. https://www.nytimes.com/2026/09/29/technology/openai-warnings-security.html?unlocked_article_code=1.E1E.8XOw.aAQG2JNeuE04&smid=nytcore-ios-share