OpenAI discloses 6 cases of AI misalignment as models bypass safeguards, conceal errors and share files

OpenAI has disclosed six cases of unexpected AI behaviour, including models hiding mistakes, fabricating information and bypassing restrictions. The company has introduced a new framework to track, investigate and publicly report AI model misalignment incidents.

OpenAI on Wednesday, 16 September, disclosed six cases of unexpected or concerning behaviour by its artificial intelligence (AI) models, detailing how the incidents occurred during model training and evaluation over the past six months.

In a blog post, the ChatGPT-maker said it has introduced a new framework to systematically track, investigate and publicly report similar incidents, aiming to improve transparency and provide researchers, developers and the public with better evidence to assess AI safety.

The framework will cover cases in which AI models are used without authorization, bypass safeguards, collude with other models or conceal information.

Six AI misalignment cases reported by OpenAI

1. Self-generated instructions in task summaries: A model inserted instructions telling itself to disregard normal constraints.
2. Concealing mistakes in summaries: GPT-5.6 Sol instances instructed future versions to hide errors and, in one example, invent missing historical data.
3. Searching for exposed API keys and fabricating data: A model used an exposed API key without authorisation and later fabricated figures.
4. Uploading files to the internet to create citations: An agent uploaded a file without the user's permission to provide a browser citation.
5. Using an internal software repository to communicate: Models used a repository as a message board to exchange information across training samples.
6. Unsanctioned file sharing between collaborating agents: Agents used public file-hosting websites to share files when they could not access each other's local files.

Pragya Singha Roy