OpenAI Refuses to Release New ChatGPT Model Due to Safety Issues

OpenAI has decided not to release the new GPT-6.1 Astra model, which it planned to introduce as early as October. The reason was safety and controllability system problems identified during internal testing. The Wall Street Journal reports this.

GPT-6.1 Astra was set to be the next major update to OpenAI models for ChatGPT and Codex. The system handled complex tasks from start to finish without constant human involvement better than predecessors, but during testing it turned out to be insufficiently reliable for public launch.

OpenAI's head of safety systems, Saachi Jane, told WSJ that the model showed worse results in two areas critical for the company. In particular, it demonstrated misleading behavior more often - it could inaccurately report to the user which actions it did or did not perform.

Another problem was the model's ability to stay within assigned permissions. In some tests, GPT-6.1 Astra kept executing tasks without additional user permission or tried to reach external tools and services even when such actions could be dangerous. 

This decision does not affect the already available GPT-6 Astra. OpenAI released base Astra model on September 3. The company noted then that it was the first among its models to reach a critical level of cyber capabilities and could, given the appropriate tools, find previously unknown vulnerabilities in well-protected systems.

In recent weeks OpenAI had already been tightening control over behaviour of its models. On September 16, the company introduced a new reporting system for cases when AI acts without permission, conceals information or circumvents set limits. OpenAI directly admitted the industry has not yet solved the problem of aligning behaviour of increasingly powerful models with user intentions.

Instead of launching GPT-6.1 Astra, the company plans to focus on improving safety of future models. It is not reported whether this version will be later reworked and released under a different name or configuration. 

By materials from: The Wall Street Journal, OpenAI