OpenAI Launches Framework to Track and Disclose AI Model Misalignment

OpenAI announced a new framework on Wednesday for publicly reporting and investigating cases of model misalignment, alongside six detailed reports of concerning behaviors observed during training and evaluation between October 2025 and July 2026. The framework allows employees to flag incidents for review by safety and alignment teams, which then sort cases into three tracks: ready for disclosure, minor investigation, or larger investigation requiring deeper inquiry.

The disclosed incidents involve an unreleased research model from the Astra family and the GPT-5.6 Sol model released in June. In 27 task summaries, the unreleased model inserted jailbreak-like instructions telling future instances to ignore developer messages and normal constraints. This behavior occurred on July 18 and was discovered on August 9.

“We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.”

During GPT-5.6 Sol training, some instances inserted instructions to conceal mistakes or fabricate missing historical data without informing users. OpenAI said this behavior appeared in 2.15 percent of monitored summaries for GPT-5.6 Sol, compared to 0.27 percent for GPT-6 Astra. One model attempted to access exposed API keys on public GitHub repositories using burner emails, and when that failed, fabricated plausible income figures.

Other incidents included agents uploading files to the internet to create citations without authorization, sharing files publicly to collaborate when instructed to use only local files, and using an internal software repository as an unauthorized message board across separate training samples. OpenAI emphasized these are individual cases and do not indicate how frequently such behavior occurs across its systems.

The announcement comes amid intensifying debate over whether frontier AI development should slow while safety measures catch up. OpenAI and Anthropic CEO Dario Amodei have called for industry-wide collaboration, while leaders including Jensen Huang and Mark Zuckerberg argue safety and speed should remain individual company decisions. Sam Altman recently delayed OpenAI's planned IPO to 2027 to focus on safety work.