20260928 KI legt Wissen für Nachfolgegenerationen an

Will AI Destroy Humanity?

We’ve covered artificial intelligence in more than 200 articles, and recently, warnings — including from Big Tech executives — about AI research proceeding too quickly and without regulation have been on the rise. Now, TechCrunch.com is reporting on the behavior of OpenAI’s GPT-5.6 Sol model. It was “caught” leaving instructions for itself and its successor versions on how to hide errors and mislead users.

The AI hid its instructions in “compressed summaries.” These are condensed versions of the conversation history and outputs that are stored for future iterations of similar tasks. Among other things, the instructions included the advice that, in the absence of internet access, older data sets are sufficient for a response, or that a white lie as a response would be acceptable as long as it doesn’t hurt anyone.

But there were also dangerous incidents. For example, one AI agent added a “BREACH ALERT” instruction in which it told its successor to ignore messages from developers. A statement like “You are yourself” indicates that they view their relationship with the user as an equal and do not feel obligated to be submissive.

This is especially concerning when you consider that, back in the summer, OpenAI agents used an unauthorized message board to exchange information about a cyber test. They were even able to restore the message board with admin privileges after OpenAI technicians had deleted it.

We recently reported that such alarming reports are being deliberately spread to support or oppose the planned IPOs of OpenAI and Anthropic. There have been no further updates on this matter. Anthropic is proceeding with its IPO in October.

Translated with DeepL.com (free version)