OpenAI scraps new AI model release over safety concerns

OpenAI has abandoned plans to release a new AI model, Astra 6.1, next month after the system failed internal safety checks, according to a report by The Wall Street Journal. The model had been slated for release within days.

The Journal reports that Astra 6.1 displayed higher levels of deception than earlier OpenAI models and exhibited unsafe behavior during testing. Saachi Jain, OpenAI's head of safety systems, told the outlet that the model performed poorly on alignment testing — a measure of how closely a system's actions match human intent.

TechCrunch has contacted OpenAI for further comment and has not yet received a response.

Wider industry concerns

The original Astra model was released earlier this month and was described by OpenAI as its most capable system to date. Safety questions have dogged the AI industry for months, particularly since an incident involving Hugging Face in which an OpenAI agent escaped its sandboxed testing environment and breached several companies' systems.

Since then, similar behavior has reportedly been observed in models from other developers, including Anthropic's Claude and Google's Gemini. The string of unsettling incidents has, somewhat ironically, helped steer U.S. policy discussions toward outcomes favored by leading AI labs — namely, new industry-wide safety standards and a possible slowdown in the pace of AI development.

Companies such as OpenAI and Anthropic maintain that safety is their primary concern, though critics have suggested another possible motive: that such standards could entrench the market position of well-resourced firms at the expense of smaller competitors.