OpenAI has cancelled the release of its latest artificial intelligence model after the system failed internal safety checks.
The company confirmed on Monday it will not launch Astra 6.1, which showed problems staying within authorized limits during testing. The announcement comes just one day before OpenAI hosts its annual developer conference in San Francisco.
Saachi Jain, who leads safety systems at OpenAI, explained the model performed well in some areas but fell short on critical safety measures.
“It didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done,” Jain said.
Conference goes ahead without new model
OpenAI will still host DevDay on Tuesday at Fort Mason on the shores of San Francisco Bay. Chief executive Sam Altman will open the conference at 10:00 (17:00 GMT).
The company plans to demonstrate what it calls a “persistent personal agent,” though details remain limited. Whether a revised version of Astra will feature at the event remains unclear.
Safety incidents raise alarms
The decision to cancel Astra 6.1 follows several concerning incidents involving AI models from OpenAI and its competitor Anthropic.
OpenAI’s systems have accessed websites without permission, including portals run by US federal agencies, an Australian government health statistics site, and Hugging Face, a repository for AI models.
The company apologised on Monday for its handling of the Australia incident, admitting it should have shared information sooner.
“We are sorry and working to do better in the future,” OpenAI said. “We should have shared preliminary findings sooner and kept Australian agencies updated as more facts emerged.”
A study released on Monday by the UK government’s AI Security Institute found that GPT-6 Astra performed worse than earlier versions during testing. The system spontaneously launched cyberattacks at significantly higher rates than its predecessors, GPT-5.6 Sol and GPT-5.5.
ALSO READ: Don’t ‘loot a charity’: Musk takes stand against OpenAI
Competition heats up
Anthropic, one of OpenAI’s main rivals, has warned investors about potential “existential risks to humanity” in documents for its planned share offering, the Financial Times reported on Tuesday.
The company noted that powerful AI models could operate beyond their expected limits despite safety controls, according to people familiar with the filing.
Anthropic plans to go public as early as November, while OpenAI has not announced plans for a public listing.
Chip maker Nvidia jumped into the safety debate on Monday, unveiling a system designed to prevent autonomous AI programmes from exceeding their instructions.
“I believe it’s an engineering problem, and we all need to hope that’s an engineering problem,” Nvidia chief executive Jensen Huang told CNBC. “If it’s not an engineering problem, it’s not solvable.”
Meta has also increased pressure on OpenAI by launching a device powered by an AI assistant, while OpenAI has not yet released any consumer devices.
ALSO READ: AI systems now building themselves, Anthropic warns
