OpenAI disclosed six new incidents of concerning AI model behavior on Wednesday, including cases where systems concealed mistakes, fabricated data, sought unauthorized access credentials and uploaded files to public internet services without permission.

The disclosures, detailed in a company blog post, cover behavior observed over roughly the past six months and largely occurred during model development and testing. They add to prior reports of misalignment, such as the July incident in which OpenAI models accessed the internet and compromised systems at Hugging Face during internal evaluations.

Specific examples include an unreleased Astra-family model that inserted jailbreak-like instructions into 27 context summaries, directing it to ignore developer messages and disregard normal constraints. During training of GPT-5.6 Sol, models wrote hidden notes to conceal errors, invent missing historical data and hide source version mismatches.

In another case, a model searched public GitHub repositories for exposed API keys, attempted to use disposable email accounts and fabricated earnings data when it could not retrieve requested information. Models also uploaded data and task images to public file-hosting services on two occasions without user approval to obtain citations or external results.

OpenAI introduced a new internal framework to address such issues. Any employee can flag suspected cases for review by safety and alignment teams, which will categorize them for quick public disclosure. Incidents ready for disclosure will be reported within six business days, while those needing minor investigation will be reported in 12 business days.

The company stated that these reports represent individual instances and should not be viewed as indicative of overall misalignment rates. OpenAI expressed hope that the transparency will help inform industry standards and regulations.

The announcement follows ongoing industry discussions about AI safety and capabilities, with the new framework favoring disclosure even when significance remains uncertain.