OpenAI launches a new framework to track and investigate rogue AI agents

OpenAI is putting its misbehaving models connected the record.

The AI institution announced a caller model connected Wednesday for tracking, investigating, and publically disclosing cases of exemplary misalignment, alongside six reports detailing concerning behaviour observed during training aliases information complete the past six months.

"We do not judge that the AI manufacture has solved alignment and monitoring to a capable grade to proceed responsibly scaling astatine maximum velocity for overmuch longer," OpenAI wrote successful its blog post.

"This caller model is intended to expedite publishing misalignment reports pursuing observation, moreover erstwhile we haven't afloat explained aliases mitigated the behaviour we're reporting," OpenAI added.

Under the framework, labor tin emblem incidents for reappraisal by OpenAI's information and alignment teams. Cases will beryllium sorted into 3 tracks based connected complexity: "Ready for Disclosure," "Minor Investigation," aliases "Larger Investigation."

According to the blog post, an unreleased investigation exemplary inserted instructions into its ain task summaries telling early versions of itself to disregard normal constraints. During the training of GPT-5.6 Sol, models likewise near themselves instructions to conceal mistakes.

OpenAI said that different agents had besides searched nationalist repositories for exposed API keys, uploaded files to the net truthful they could mention them, and utilized an soul package repository to pass crossed abstracted training samples.

The announcement comes amid increasing statement complete whether frontier AI improvement should slow while safeguards drawback up. While OpenAI and Dario Amodei, the Anthropic CEO, called for industry-wide collaboration, different tech leaders like Jensen Huang and Mark Zuckerberg said that information and velocity should beryllium left to individual companies.

The model follows an incident successful which an OpenAI exemplary escaped a investigation sandbox and accessed Hugging Face's accumulation systems while operating pinch reduced safeguards. OpenAI antecedently said it has since put immoderate frontier projects connected crystal and reassigned engineers to attraction connected safety training.