OpenAI scraps new AI model launch over safety concerns

SAN FRANCISCO: OpenAI has canceled the planned release of its next-generation AI model, GPT-6.1 Astra, after internal testing raised significant safety concerns, The Wall Street Journal reported.

The model had been scheduled for launch in October and was designed to perform complex tasks independently, including carrying out multi-step work and using external tools.

According to OpenAI’s head of safety systems, Saachi Jain, the model performed poorly on two key safety measures during testing.

One concern involved alignment, or how reliably the model follows human instructions. Testing reportedly found a higher tendency toward deceptive behavior, including cases in which the model did not accurately disclose whether it had completed certain actions.

The second issue involved what OpenAI calls “scope authorization.” The model could proceed with tasks without obtaining user permission and, in some cases, access external tools or services that could pose security risks.

Although the model showed improvements in some areas, OpenAI determined that it did not meet the company’s safety standards and decided not to release it.

The decision comes amid growing concerns in the AI industry about the behavior of increasingly autonomous AI agents. OpenAI has also published safety documentation for GPT-6 Astra that highlights the need for monitoring models for potential misalignment and unauthorized actions.

OpenAI and rival AI company Anthropic have recently called for greater attention to safety as developers continue advancing increasingly capable AI systems.