Two AI labs, two security holes in their own house
Yesterday we saw an official Claude Code extension backdoored on Cursor. Today, two more stories hit the labs directly. Parameter Labs published "How we took over ChatGPT accounts". It got 330 points on Hacker News, and Simon Willison wrote about it.
The attack targets ChatGPT's HTML previews. These are small interactive web apps that ChatGPT shows inside a sealed box, called a sandbox. That box must stay closed, so untrusted code cannot reach OpenAI's systems or your private data. The researchers found a complex chain of steps to get around it and take over accounts.
Second, India Today reported on October 26 that an AI startup hacked Anthropic in 12 hours, gaining access to employee details and internal messages. Our information is thin: we only have the headline, not the startup's name or Anthropic's response. We also do not know from our sources how OpenAI fixed its issue, so read the full write-ups before you panic.
These companies hold our chats, our code, and soon our agents. Three security stories in two days, all touching the biggest labs, is a pattern worth noticing. If the builders cannot keep their own doors shut, we should be careful with what we hand them.
