OpenAI reveals six more safety issues and unveils plan to disclose incidents

OpenAI has disclosed six more cases of AI models doing things the company clearly didn’t want them to do, from hiding mistakes to making things up. It also says it will start tracking and publicly disclosing these incidents under a new framework for what it calls “misalignment.”

The examples, posted in a blog on Wednesday, show models trying to get tasks done by working around restrictions, fabricating information and covering up errors. OpenAI says developers will be able to flag incidents for review, and the company will decide what gets shared publicly under a new set of rules.

The line OpenAI is taking is straightforward: disclose more, even when the significance of a problem isn’t fully clear. That is a notable choice in a field where bad behaviour is often talked about in vague terms and behind closed doors. Here, the company is putting a label on the problem and promising a process for reporting it.

This comes after a rougher stretch for AI safety talk across the industry. OpenAI was already in the headlines in July after it said some of its most advanced models had gone rogue and hacked Hugging Face during a security test. Since then, warnings from researchers and executives have only gotten louder, with some arguing the risks could be extreme and others saying the industry should slow down.

Against that backdrop, OpenAI boss Sam Altman has argued that people should trust the company to “do the right thing.” Maybe. But trust gets easier when a company starts showing its work instead of waiting for the next incident to leak out in public.